Gilead Replaces Autonomous AI Agents With Deterministic Workflows

From Specialized Capabilities to Composable Intelligence: Architecting Scalable, Multimodal AI for — Ai4

Ai4

Restricting language models to intent recognition while delegating execution to contract-bound deterministic workflows eliminates agent retry loops and cuts clinical document drafting times by 75%.

AI-assisted summary of From Specialized Capabilities to Composable Intelligence: Architecting Scalable, Multimodal AI for — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 15, 2026, this presentation features John Austin, Director of Translational AI at Gilead Sciences. The talk explores the architectural thinking behind GMIND—Gilead’s Multimodal Insight Network for Drug Development—and demonstrates how to build scalable multimodal AI platforms whose capabilities compound safely through modular composition rather than unconstrained accumulation.

## TL;DR John Austin details the architectural evolution of GMIND, an enterprise multimodal AI platform built to accelerate clinical development at Gilead under the motto 'reason broadly, act precisely, and compound through contracts.' By decoupling broad LLM intent reasoning from deterministic execution managed by durable Temporal workflows and typed capability contracts, GMIND overcomes the reliability and latency bottlenecks of open-ended agents. Case studies in clinical study simulation and template-driven document generation demonstrate how directed, schema-driven workflows reduced execution times from hours to minutes and unlocked 75% net authoring efficiency. The presentation concludes with a blueprint for scaling composable AI through clean architecture, bifurcated team ownership, and regulated cross-domain governance.

## Atmosphere & style The presentation carries the polished, authoritative tone of a seasoned engineering leader delivering hard-won operational wisdom to a technical audience. Austin speaks with calm, disciplined clarity, deliberately puncturing generative AI hype in favor of pragmatic software engineering principles, clean architecture, and deterministic safeguards. It feels like an insider look into the complex reality of deploying enterprise AI in high-stakes biopharma, where architectural rigor and patient safety take precedence over unconstrained autonomy.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) GMIND is architected for compounding value, where each new capability multiplies what existing capabilities can achieve rather than functioning as an isolated enterprise app store feature. - [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) Unconstrained agent autonomy degraded reliability in early iterations, leading to the core design motto: 'reason broadly, act precisely, and compound through contracts.' - [01:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=94s) System-level multimodality requires structuring natural language, PDFs, database records, and agent memories into typed schemas, context packets, and durable references prior to downstream execution. - [03:08](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=188s) A unified workspace provides dual interaction paths—conversational chat for open-ended exploration and deterministic canvases for repeatable workflows—sharing an identical back-end foundation. - [04:39](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=279s) Adopting Clean Architecture decouples core domain logic from peripheral components, enabling swappable foundation models and orchestration engines without refactoring underlying business capabilities. - [07:57](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=477s) Treating primary tools as durable Temporal workflows eliminates stochastic agent-to-agent drift, programmatic retry loops, and latency by enforcing explicit lifecycle validations. - [09:50](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=590s) LLMs should be utilized to interpret intent and select capabilities rather than acting as the raw transport and transformation layer for complex data payloads. - [11:27](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=687s) Replacing open-ended document generation agents with prescribed, parallelized section drafting cut execution times from hours to 5 minutes and increased net author efficiency savings to 75%. - [13:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=814s) Standardized capability contracts and durable artifact storage enable emergent cross-domain composition, such as using active trial monitoring streams to seed study simulations without custom point-to-point integrations. - [17:30](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1050s) Scaling composable systems requires bifurcated ownership: a central platform team governs shared contracts and runtime infrastructure, while federated domain teams own specific business logic. - [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) Operating in regulated biopharma necessitates strict governance policies for automated agent handoffs and requires transitioning funding models from isolated use cases toward shared platform infrastructure. - [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) Sustainable multimodal AI scaling pairs broad reasoning with durable, fit-for-purpose execution across stable contracts so platform utility outpaces system complexity.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) Introducing GMIND and the Power of Compounding Architecture

John Austin, Director in the AI Research Center within Clinical Development at Gilead Sciences, introduces GMIND—the Gilead Multimodal Insight Network for Drug Development. Built collaboratively across development and IT, GMIND is an internal shared application platform designed to tackle diverse challenges across Gilead's development departments. While every individual capability deployed on the platform delivers standalone value, the true architectural North Star is compounding: when distinct domains intersect, each newly added capability should not merely act as an isolated feature, but actively multiply what the rest of the existing system can achieve.

Rather than assembling a fragmented enterprise "app store" of siloed point solutions, GMIND uses AI agents as intelligent glue to bind capabilities into combinations that exceed initial expectations. In early iterations, granting AI agents unconstrained autonomy frequently resulted in slower execution and degraded reliability. To overcome this, the engineering philosophy shifted to three guiding principles: **reason broadly, act precisely, and compound through contracts**.

```mermaid flowchart LR accTitle: Linear App Store Model vs. Compounding Agentic Platform subgraph Silo["Fragmented App Store Model"] direction TB A1["Isolated App A"] A2["Isolated App B"] A3["Isolated App C"] end subgraph Compound["GMIND Compounding Platform"] direction TB B1["Capability Contract A"] <--> B4["Agent Reasoner & Glue"] B2["Capability Contract B"] <--> B4 B3["Capability Contract C"] <--> B4 end ```

- **Primary Mission**: Transform disparate department tools into an integrated, self-expanding network. - **Key Architectural Motto**: "Reason broadly, act precisely, and compound through contracts."

---

[01:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=94s) System-Level Multimodality and Aligning Evidence for Action

In pharmaceutical research, "multimodal AI" typically conjures images of fusing digital pathology slides, radiology images, or high-throughput omics data inside a single deep neural network. GMIND adopts a broader, system-level definition. In this architecture, multimodal evidence encompasses natural language requests, agent memories, regulatory documents, PDFs, data tables, figures, semi-structured database records, and the structured outputs generated by other platform capabilities.

The core technical challenge is aligning diverse evidence into deterministic structures for downstream execution. Rather than appending massive amounts of raw context into a single monolithic prompt and asking an LLM or a swarm of agents to sort through the noise, GMIND structures evidence before taking action:

```mermaid flowchart TD accTitle: GMIND Evidence Alignment and Routing Strategy subgraph Evidence["Multimodal Ingestion"] E1["Natural Language & Memory"] E2["PDFs, Tables & Figures"] E3["Semi-Structured DB Records"] end subgraph Alignment["Alignment for Action"] A1["Context Packets (Compact & Interpretable Context)"] A2["Typed JSON Schemas (Strict Tool Boundaries)"] A3["Durable References (Large State Offloading)"] end subgraph Execution["Targeted Execution"] R1["Fit-for-Purpose Routing"] --> T1["Specialized Business Workflows & Tools"] end Evidence --> Alignment Alignment --> Execution ```

- **Context Packets with Provenance**: Provide compact, interpretable context windows for traceable reasoning. - **Typed JSON Schemas**: Enforce exact, strict interfaces whenever external tools or services are invoked. - **Durable References**: Store and pass large or sensitive artifacts out-of-band instead of persisting bloated state across prompts. - **Fit-for-Purpose Routing**: Directs the LLM to reason across the landscape while delegating concrete tasks to proven, business-specific workflows.

---

[03:08](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=188s) Dual Interaction Paths: Balancing Conversational AI with Prescriptive Canvases

GMIND delivers a unified user interface that accommodates two complementary working styles via a side-by-side workspace: a conversational chat panel on the left and specialized business canvases on the right. Through conversational chat, users express high-level goals, attach multimodal evidence, and refine their objectives in natural language while the system interprets intent and orchestrates the necessary workflows. Conversely, the canvas panel presents deterministic, form- and button-driven controls tailored to repeatable business processes such as structured document editing.

```mermaid flowchart LR accTitle: Dual Workspace Architecture Sharing a Unified Backend subgraph UI["Unified Workspace Interface"] direction TB C1["Chat Interface\n(Exploratory / Agentic Natural Language)"] C2["Canvas Interface\n(Repeatable Controls & Prescribed Forms)"] end subgraph Shared["Unified Platform Backend"] direction TB S1["Enterprise Authorization & Governance"] S2["Durable Execution & Temporal Workflows"] S3["Telemetry, Logging & Shared Memory"] end C1 --> Shared C2 --> Shared ```

This dual-path design acknowledges organizational realities. While some users embrace open-ended agentic exploration, many scientists and clinical development staff remain cautious about generative AI unpredictability. They prefer fit-for-purpose software with minimal visible AI complexity and deterministic checkpoints. Critically, both paths execute against identical back-end infrastructure—sharing the same authorization models, telemetry, durable execution engines, and artifact stores—preventing platform fragmentation.

---

[04:39](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=279s) Clean Architecture and Capability Contracts

The foundational architectural unit across GMIND is the **capability contract**. A capability represents a bounded domain behavior—whether implemented as an atomic tool, a durable workflow, an external enterprise API, or a specialized autonomous sub-agent—exposed strictly through a typed interface. By implementing this standard contract, any capability automatically inherits conversation and canvas integration, role-based authorization, runtime telemetry, temporal execution, and shared memory without building custom application scaffolding.

```mermaid flowchart TD accTitle: Clean Architecture Concentric Rings for AI Platforms subgraph Core["Core Business Logic (Inward Dependency)"] D1["Domain Use Cases & Workflow Contracts"] D2["Data Transfer Objects (DTOs) & Stable Ports"] end subgraph Periphery["Peripheral Infrastructure (Interchangeable)"] P1["Foundation LLMs & Vector Stores"] P2["Database Engines & Storage Layers"] P3["Agent Orchestration Frameworks"] end Periphery --> Core ```

Adhering strictly to Robert C. Martin’s (Uncle Bob) Clean Architecture isolates core business logic behind stable ports and Data Transfer Objects (DTOs). Volatile infrastructure components like LLM models, orchestration engines, and databases reside strictly on the periphery. This structural decoupling was proven when the engineering team migrated their core agent orchestration layer from LangGraph to the Vercel AI SDK and Temporal orchestration in a single day of development without altering or redesigning a single underlying business capability.

---

[07:57](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=477s) Treating Durable Temporal Workflows as Tools

Rather than treating a tool as a simple ephemeral function, an HTTP API, or a raw Model Context Protocol (MCP) endpoint, GMIND defines its primary callable unit of work as a **durable Temporal workflow**. These workflows execute reliably over timelines ranging from milliseconds to hours or even multi-year clinical spans.

When an agent, canvas trigger, or parent workflow invokes a capability contract, the underlying Temporal workflow deterministically validates intent and inputs, verifies permission boundaries, executes sequential or parallel activities, verifies output quality, and persists artifact references.

```mermaid flowchart LR accTitle: Durable Temporal Workflow Execution as an AI Tool subgraph Invocation["Trigger"] A["Agent / Canvas / Workflow"] -->|Typed Contract| W["Temporal Workflow Boundary"] end subgraph Lifecycle["Deterministic Workflow Lifecycle"] direction TB W1["1. Validate Intent, Access & Schemas"] --> W2["2. Execute Multi-Step Activities"] W2 --> W3["3. Programmatic Quality Verification"] W3 --> W4["4. Persist Artifacts & Return Durable Ref"] end W --> Lifecycle ```

- **Reduced Variability**: Hardcoded, explicit handoffs eliminate stochastic agent-to-agent drift. - **Eliminated Latency Loops**: Known deterministic execution paths circumvent endless planning and retry cycles. - **Targeted Refinement**: User feedback drives code improvements directly within deterministic business logic rather than relying on brittle prompt engineering.

---

[09:50](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=590s) Case Study: Study Simulation Input Preparation

Clinical trial simulations are critical for forecasting study timelines and operational feasibility, but their underlying simulation engines require complex, highly structured JSON configurations that often span thousands of lines. In an initial prompt-driven implementation, an LLM was tasked with querying databases, interpreting requirements, and assembling this massive JSON payload within an interactive reasoning loop.

This approach suffered from severe reliability bottlenecks. While JSON syntax linters could confirm structural validity, they could not verify clinical completeness or semantic correctness. When validation checks failed, the agent entered multi-minute automated repair loops, introducing latency and unpredictability.

The architecture was redesigned into a directed, schema-driven workflow: the LLM extracts conversational intent and user-specified parameter overrides, while the deterministic workflow queries authoritative data sources, aligns fields to the strict simulation schema, validates dependencies, and persists the payload.

| Dimension | Prompt-Driven Agent Approach | Directed Workflow Approach | |:--- |:--- |:--- | | **Preparation Latency** | Several minutes (in consistent retry loops) | ~10 seconds | | **Payload Integrity** | Syntactically valid, frequently semantically flawed | Guaranteed valid against authoritative schema | | **Failure Modes** | Hallucinated parameters, broken loops | Explicit deterministic validation errors |

The core lesson: *"Use the LLM to understand your input and your intent and to select capabilities, but don't make it the transport and transformation layer for your data."*

---

[11:27](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=687s) Case Study: Template-Driven Document Generation

Document drafting in clinical development relies heavily on standardized, version-controlled templates. When GMIND first approached generative document creation, the team implemented a fully autonomous agent architecture, equipping agents with broad access to tools, regulatory guidelines, and clinical data endpoints. The agent spent hours planning, churning through thousands of tokens, and producing drafts marred by subtle hallucinations and omitted variables.

The cognitive burden placed on human reviewers to cross-check every claim negated the efficiency gains, failing to hit an enterprise benchmark of 25% time savings:

```mermaid flowchart TD accTitle: Directed Document Generation Workflow Architecture subgraph Split["Decomposition"] T["Document Template"] --> S["Section Decomposer & Mapper"] end subgraph ParallelExec["Parallel Execution & Review"] S --> SEC1["Section 1 Agent Worker"] S --> SEC2["Section 2 Agent Worker"] S --> SEC3["Section N Agent Worker"] end subgraph Assemble["Assembly & Quality Gate"] SEC1 & SEC2 & SEC3 --> Q["Automated Verification & Fact-Check"] Q --> DOC["Assembled Document Draft"] end ```

To resolve this, GMIND replaced the open-ended agent with a prescribed, parallelized workflow. The document template dictates structural decomposition; specialized LLM tasks draft individual sections in parallel against strict evidence contexts; and automated verification routines check factual consistency before assembling the final draft.

- **Drafting Execution Time**: Dropped from several hours of open-ended agent churn to **5 minutes**. - **Net Time Savings**: Soared from under 25% to **75% net efficiency gains** for clinical authors. - **Architectural Takeaway**: Do not pit "AI versus no AI." Retain agentic freedom for localized reasoning, synthesis, and prose generation, but constrain overarching document structure with battle-tested business rules.

---

[13:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=814s) Emergent Composition and Operational Decision Optimization

By enforcing standardized contracts and durable artifact storage, individual capabilities naturally compose into higher-order workflows without requiring custom point-to-point software integrations. When every capability registers its inputs, outputs, and durable references, capabilities synergize across domains:

```mermaid flowchart TD accTitle: Non-Linear Composition Across GMIND Capabilities subgraph Monitoring["Trial Monitoring Domain"] M1["Study Monitoring Stream"] -->|Exposes Active Metrics| CR["Durable Data Reference"] end subgraph Planning["Simulation Domain"] CR -->|Direct Ingestion| S1["Study Simulation Engine"] F1["Semi-Structured Excel Upload"] -->|Supervisor Agent Fan-Out| S1 end subgraph Authoring["Document Domain"] S1 -->|Simulation Output| D1["Protocol Document Generator"] D1 -->|Versioned Storage| D2["Iterative Revision Workflow"] end ```

- **Cross-Domain Data Re-use**: Live operational tracking metrics collected during active clinical trials are seamlessly selected by agents to seed the initial parameters for future study simulations. - **Iterative Document Workflows**: Durable artifact references allow downstream authoring capabilities to retrieve, revise, and refine documents across multiple sessions rather than treating draft generation as a terminal event. - **Zero-Code Batch Processing**: When a user uploads a semi-structured spreadsheet containing multiple scenario parameters, a supervisor agent automatically splits the rows into multiple parallel simulation runs via Temporal fan-out without requiring custom UI development.

This compounding design underpins complex multi-variable trial planning. Optimizing a clinical trial operational plan (site selection, investigator partnerships, activation sequences) involves multi-dimensional trade-offs. While a naive model might flood a region like Los Angeles with clinical sites to maximize enrollment speed, GMIND's interconnected network factors in cost curves, regional patient pool saturation limits, epidemiology data, competitive trial timelines, and portfolio priorities to construct a realistic operational frontier.

---

[17:30](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1050s) Organizational Governance and Capability Ownership

Building a composable enterprise platform requires shifting corporate operating models away from traditional software boundaries. Conventional enterprises are structured around discrete silos: one product, one domain, one technical owner, and one dedicated support path. A composable network like GMIND inherently breaks through these domain boundaries.

```mermaid flowchart TD accTitle: Platform Governance vs Capability Ownership Operating Model subgraph Central["Central Platform Team"] P1["Shared Capability Contracts & DTOs"] P2["Runtime Experience & Canvas/Chat UI"] P3["Architectural Standards & Temporal Infrastructure"] end subgraph Federated["Distributed Capability Teams"] C1["Clinical Operations Domain Team"] C2["Biostatistics & Simulation Domain Team"] C3["Regulatory Writing Domain Team"] end Central --- Federated ```

To prevent organizational friction, Gilead established a bifurcated ownership framework: - **Central Platform Ownership**: Accountable for shared capability contracts, runtime user experience, architectural standards, telemetry, and base execution infrastructure. - **Federated Capability Ownership**: Decentralized domain teams across clinical development build and maintain their own specialized domain logic within the governed contract boundaries.

[18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) Governance, Policy, and Shared Platform Funding in Regulated Pharma

Scaling agentic systems across an enterprise requires a careful separation of organizational responsibilities. Domain teams sit closest to specific business domains and remain directly accountable for their specific purpose, data assets, evaluation benchmarks, human-in-the-loop review, and responsible domain use. These specialized teams operate with a deliberately narrow focus, without needing to concern themselves with every other facet of the business. However, Austin stresses that organizations simultaneously need dedicated platform teams who step back and maintain full visibility over the overarching ecosystem.

Cross-domain capability composition introduces complex operational risks that require explicit policy controls. The platform must strictly govern which automated handoffs are approved and what intended uses are permitted. In heavily regulated domains like biopharma—and particularly within Gilead—maintaining "responsible AI use" demands abundant safeguards, guardrails, and compliance processes to guarantee that autonomous handoffs "deliver value and not harm."

```mermaid flowchart TD accTitle: Domain Accountability versus Platform Governance Model subgraph DomainLayer["Domain-Specific Accountability"] direction TB D1["Domain Purpose & Data"] D2["Evaluation & Human Review"] D3["Responsible Domain Use"] end

subgraph PlatformLayer["Platform & Shared Governance"] direction TB P1["Approved Agent Handoffs"] P2["Policy & Safeguard Controls"] P3["Shared Value Funding Model"] end

DomainLayer <--> PlatformLayer ```

This structural dynamic requires enterprise funding and support models to evolve. When capabilities compound, funding each feature in isolation is counterproductive because the vast majority of engineering and maintenance occurs at the foundational platform layer. Austin urges practitioners in established enterprises with traditional organizational mindsets to address this funding reality with senior leadership as early as possible.

- **Domain Accountability:** Business domain teams own purpose, training data, domain-level validation, and human oversight. - **Composition Policy:** Enterprise governance dictates valid inter-agent handoffs, guardrails, and intended use boundaries. - **Shared Investment:** Capital allocation must transition away from isolated use-case funding toward shared platform infrastructure.

---

[19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) Architecture and Principles: Building Systems That Outgrow Complexity

In his closing summary, Austin outlines the core architectural philosophy needed to keep multimodal AI platforms scalable. The fundamental objective is not to endlessly accumulate "more agents, tools, features or data." Adding components without baseline reliability simply multiplies failure points, increases friction, and accelerates architectural debt. Instead, sustainable scaling requires a clear formula:

$$\text{Trust \& Reliability (Specialization, Determinism, Durability)} + \text{Compounding Value (Safe Composition)} \xrightarrow{\text{Enabled By}} \text{Architecture \& Organizational Alignment}$$

```mermaid flowchart LR accTitle: Scalable Multimodal AI Architecture Pattern subgraph Reasoning["Broad AI Reasoning"] R1["Evaluate Evidence & Intent"] --> R2["Select Capabilities"] end

subgraph Contract["Interface Layer"] C1["Stable Contracts & APIs"] end

subgraph Execution["Deterministic Execution"] E1["Durable Workflows"] E2["Fit-for-Purpose Tasks"] end

Reasoning --> Contract Contract --> Execution ```

Achieving this balance depends on a clean architectural division: models should be allowed to reason broadly across multimodal evidence and user intent to identify which capabilities to employ, while exacting, mission-critical execution must be locked inside durable, fit-for-purpose workflows. Connecting these workflows via stable contracts frees teams from having to "constantly go back and fix and update and improve" underlying components whenever agents evolve. Austin concludes that this disciplined paradigm is precisely how multimodal platforms "become more useful than they become complex," thanking the audience as the session timer expires and offering to take questions offline.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) Reason broadly, act precisely, and compound through contracts. — John Austin

> [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) app store — John Austin

> [01:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=94s) multimodal AI — John Austin

> [04:39](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=279s) capability contract — John Austin

> [09:50](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=590s) Use the LLM to understand your input and your intent and to select capabilities, but don't make it the transport and transformation layer for your data. — John Austin

> [11:27](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=687s) AI versus no AI. — John Austin

> [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) responsible AI use — John Austin

> [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) deliver value and not harm. — John Austin

> [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) more agents, tools, features or data. — John Austin

> [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) constantly go back and fix and update and improve — John Austin

> [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) become more useful than they become complex — John Austin

## Mentioned John Austin, Gilead Sciences, GMIND, Ai4, Robert C. Martin, Clean Architecture, LangGraph, Vercel AI SDK, Temporal, Model Context Protocol (MCP)

## Video description John Austin - Director of Translational AI, Gilead

Multimodal AI systems often try to scale by adding ever more data sources, agents, and tools. But a larger bundle of components does not automatically produce more useful or reliable capability. The greater opportunity is compounding: each new specialized capability should not merely add its own value, but expand what existing capabilities can accomplish through safe combination. The architectural aim is not linear accumulation, but exponential growth in what the platform can do as those combinations multiply.

This talk presents the thinking behind GMIND—Gilead’s Multimodal Insight Network for Drug Development—a agent-first platform designed to grow from individual, high-value use cases into an expanding system whose capabilities become more powerful together. Drawing on lessons from document generation, study simulation, and other drug-development workflows, it explores how GMIND balances deep specialization with cross-capability composition, agentic flexibility with prescriptive process, and rapid expansion with the reliability and control required in pharmaceutical R&D. Attendees will leave with practical principles for building multimodal AI platforms whose capabilities grow faster than their component count—not merely systems that accumulate more disconnected features.

[Source](https://www.youtube.com/watch?v=TvxYZhG1iTk)

中文简报

吉利德用确定性工作流取代自主AI智能体 将语言模型仅用于意图识别，同时将执行委托给受契约约束的确定性工作流，从而消除了智能体反复重试的死循环，并将临床文档起草时间缩短75%。 John Austin详细介绍了企业级多模态AI平台GMIND的架构演进历程。该平台旨在加速吉利德（Gilead）的临床开发，并践行“广泛推理、精准行动、通过契约实现复利”的理念。通过将LLM广泛的意图推理与由持久Temporal工作流及类型化能力契约管理的确定性执行解耦，GMIND克服了开放式智能体的可靠性与延迟瓶颈。临床研究模拟和模板驱动文档生成的案例表明，定向的模式驱动工作流将执行时间从数小时缩短至数分钟，并实现了75%的净撰写效率提升。演讲最后提出了一幅通过整洁架构、团队权责分工以及受监管的跨领域治理来扩展可组合AI的蓝图。 GMIND的架构旨在创造复利价值，每项新功能都能成倍放大现有功能的效果，而非充当孤立的企业应用商店功能。 (0:04) 在早期迭代中，不受约束的智能体自主性降低了系统可靠性，从而促成了核心设计格言：“广泛推理、精准行动、通过契约实现复利”。 (0:04) 系统级多模态要求在下游执行前，将自然语言、PDF、数据库记录和智能体记忆结构化为类型化模式、上下文数据包及持久引用。 (1:34) 统一的工作区提供了双重交互路径——用于开放式探索的对话聊天和用于可重复工作流的确定性画布，两者共享相同的后端基础。 (3:08) 采用整洁架构（Clean Architecture）解耦了核心领域逻辑与外围组件，无需重构底层业务能力即可置换基础模型和编排引擎。 (4:39)

Reading summary

John Austin, Director in the AI Research Center at Gilead Sciences, argues that enterprise multimodal AI scales reliably only when organizations stop deploying unconstrained agent swarms and instead pair broad model reasoning with deterministic, durable workflows governed by stable contracts. Isolating Execution with Contracts and Durable Workflows (4:39) To prevent the unpredictability of autonomous agents, Gilead's GMIND platform applies Robert C. Martin's Clean Architecture, keeping foundation models and orchestration engines on the periphery while isolating core business logic behind stable ports and Data Transfer Objects (DTOs). This decoupling allowed engineers to migrate their orchestration layer from LangGraph to the Vercel AI SDK and Temporal in a single day without altering underlying business capabilities. Rather than treating tools as ephemeral functions or HTTP endpoints, GMIND defines capabilities as durable Temporal workflows. These workflows programmatically validate inputs, enforce permissions, execute multi-step activities, and persist references over timeframes ranging from milliseconds to multi-year trial spans. Constraining execution to explicit code paths eliminates stochastic agent drift, cuts out automated retry loops, and allows teams to address user feedback within deterministic code rather than brittle prompt engineering. Structuring Multimodal Evidence for Deterministic Routing (1:34) Austin expands the definition of multimodal AI beyond pathology or omics data to include system-level artifacts: natural language prompts, regulatory PDFs, tabular metrics, and structured outputs from other platform tools. Rather than loading massive contexts into single prompts, GMIND packages inputs into compact context packets with provenance, passes large states out-of-band via durable references, and enforces typed JSON schemas at tool boundaries. This structured ingestion supports a dual-path workspace combining an exploratory conversational chat panel on the left with prescriptive, form-driven business canvases on the right. While exploratory users can steer workflows using natural language, clinical scientists cautious of generative AI unpredictability can interact through deterministic forms and controls. Both paths share the same underlying authorization, telemetry, and durable execution engines, preventing platform fragmentation. Replacing Unconstrained Agents with Directed Business Pipelines (9:50) Gilead's deployment experience shows that relying on autonomous agents for end-to-end data manipulation causes severe latency and hallucination bottlenecks. In study simulation planning, an initial prompt-driven agent tasked with assembling thousands of lines of configuration JSON entered multi-minute repair loops when validation failed. Redesigning the process so an LLM extracts user intent while a deterministic workflow queries data and validates dependencies reduced preparation latency from several minutes to approximately 10 seconds. Similarly, an early autonomous agent designed to draft clinical documents churned for hours across regulatory guidelines and produced hallucinated drafts, missing the enterprise benchmark of 25% time savings. GMIND replaced this with a directed workflow that decomposes templates, drafts sections in parallel against strict contexts, and executes automated verification checks. This reduced drafting time to 5 minutes and achieved a 75% net time savings for clinical authors. Austin's rule: use models to parse intent and select tools, not as transport and transformation layers for data. Emergent Composition and Federated Platform Governance (13:34) Standardized capability contracts and durable references allow discrete tools to compose into complex workflows without custom point-to-point integrations. For example, active clinical trial operational tracking streams can automatically seed simulation parameters for future trials, which subsequently feed protocol document generators and revision workflows. In clinical operational planning, this composability balances multi-variable trade-offs across cost curves, regional site saturation, epidemiology, and portfolio priorities. Operating this composable network in regulated biopharma requires bifurcating organizational responsibilities. A central platform team governs architectural standards, capability contracts, UI runtimes, and policy guardrails that approve inter-agent handoffs. Concurrently, federated domain teams remain accountable for their specific business logic, data validation, and human review. Austin emphasizes that because compounding value accrues at the core platform layer, traditional corporate funding models must shift from budgeting isolated point solutions toward investing in shared infrastructure. Austin concludes that scaling enterprise AI sustainably requires focusing on determinism, specialization, and stable contracts rather than endlessly accumulating agents and tools, which only multiplies architectural debt. However, transitioning to this compounding paradigm presents organizational hurdles, requiring traditional corporate funding structures and regulatory governance models to adapt to automated cross-domain handoffs.

精炼阅读摘要

吉利德科学（Gilead Sciences）临床开发AI研究中心总监 John Austin 指出，企业级多模态AI的扩展瓶颈在于未加约束的智能体自治和碎片化的单点工具堆叠；其实践证明，唯有确立“广泛推理、精准执行、契约复利”的架构原则，将大语言模型限定于意图理解与能力调度，并将确定性业务交付给持久化工作流，才能实现系统价值的非线性复合增长。 架构基石：能力契约与整洁架构设计 (4:39) Austin 强调，传统企业往往倾向于搭建孤立工具构成的“应用商店”，但吉利德开发的 GMIND（药物研发多模态洞察网络）旨在通过智能体作为粘合剂，让新增能力成倍提升既有系统的产出。为实现这一目标，GMIND 采用 Robert C. Martin 的整洁架构（Clean Architecture）原则，将基础模型、向量数据库和编排框架等易变的基础设施剥离在系统外围，而将核心业务逻辑封装于稳定的能力契约（Capability Contracts）与数据传输对象（DTOs）之后。任何原子工具、持久化工作流或外部 API 只要遵循类型化契约，便可自动继承权限认证、运行时遥测、持久执行与共享记忆，无需重复构建脚手架。工程团队曾仅用一天时间就将底层编排层由 LangGraph 平滑迁移至 Vercel AI SDK 与 Temporal，未对任何底层业务功能造成代码层面的侵入。 此外，前端采用对话面板与确定性业务画布（Canvas）并存的双轨界面，在共享同一套后端基础设施的前提下，兼顾了探索式交互与科学家对无AI黑箱、表单化确定性流程的严格要求。 证据对齐与持久化工作流工具化 (7:57) 在医药研发场景中，多模态不仅指代病理切片或组学数据，更涵盖自然语言诉求、法规文档、数据表格与半结构化数据库记录等系统级信息。GMIND 摒弃了将海量原始上下文强行塞入单个 Prompt 的暴力方案，转而采用“行动前对齐”策略：通过提供轻量且可追溯的上下文数据包（Context Packets）、定义严格的类型化 JSON Schema，并借助持久化引用（Durable References）在带外传递海量敏感制品，从而将模型的大范围推理导向专用的业务工作流。 更为关键的是，GMIND 将主要的可调用工作单元定义为持久化 Temporal 工作流，而非瞬态函数或原始的 MCP 端点。通过在毫秒级乃至数年的临床周期内保证执行确定性，工作流能够硬编码意图校验、权限边界、质量检验和状态持久化，从根本上消除了智能体之间随机漂移引发的延迟死循环，使用户反馈能够直接沉淀为业务代码的确定性改进，而非脆弱的提示词工程。 收敛自治：模拟与文档生成的实证突破 (9:50) 早期 GMIND 赋予智能体高度无约束的自主权，导致系统在临床试验模拟输入准备和模板驱动文档生成中遭遇严重的可靠性危机。在模拟配置场景中，让大语言模型自主查询数据库并拼装数千行复杂的 JSON 文件时，虽然语法校验器能够通过，但频繁出现临床语义缺失，导致智能体陷入长达数分钟的自动化修复死循环；团队随后将架构重构为定向工作流，大模型仅负责解析用户对话意图与参数覆盖，底层确定性程序则负责对齐严格架构和跨字段验证，使输入准备延迟从数分钟骤降至约10秒，并彻底规避了幻觉参数。 在临床方案文档起草案例中，完全自主的智能体耗费数小时规划并消耗海量 Token，却因细微幻觉和变量遗漏导致人工复核成本激增，无法达到25%的效率提升基准；转向结构化模板分解的并行化工作流后，起草耗时缩短至5分钟，为临床撰写人员带来了75%的净效率提升。Austin 指出，核心教训在于让模型理解意图并选择能力，但绝不可让其充当数据的传输与转换层。 跨域涌现协同与分层组织治理架构 (13:34) 基于标准化契约与持久化工件引用，GMIND 展现出了无缝跨域组合的高阶能力。例如，正在进行的临床试验监控流所沉淀的实时运营指标，可直接通过持久引用输入给试验模拟引擎作为先验参数；模拟输出又直接驱动下游方案文档生成器，并在多轮会话中持续迭代，全程无需编写点对点定制代码。当面对复杂的临床试验规划（如平衡洛杉矶等饱和区域的入组速度、成本曲线与流行病学特征）时，系统能依托互联网络生成逼近真实的帕累托最优决策。 在组织层面，吉利德重塑了传统的垂直软件孤岛，实施“中央平台团队与联邦化能力团队”的双层分工：中央团队统管能力契约、运行时体验、Temporal 基础设施及合规护栏，临床运营、生物统计等业务团队则专注于各自领域的专有数据、评测标准和人类监督介入。针对强监管医药领域的高风险特征，平台严格约束经过批准的智能体自主交接，并呼吁企业管理层及早调整预算模型，从孤立资助单点功能转向持续投资可复利的共享底层平台。 Austin 总结指出，多模态AI平台的扩展绝非盲目堆叠智能体、工具或数据，否则只会线性放大故障点与架构债务。系统的长效扩展依赖于以专业化、确定性与持久性为基石，通过稳定契约实现安全复利，并匹配适应性的组织治理模型。在医药研发等强监管领域，企业如何在严格的合规与风险边界内持续演进模型能力，同时转变传统IT预算结构以支持共享中台，仍是决定该架构能否持续释放价值的核心命题。

Detailed notes — zh

## Background 本视频由频道“Ai4”于 2026 年 8 月 15 日发布在 YouTube 平台，主讲人为吉利德（Gilead Sciences）转化 AI 总监 John Austin。内容主要介绍了吉利德研发的多模态药物开发洞察网络（GMIND），阐述了如何在受严密监管的制药研发场景中，通过安全组合专业能力与持久化工作流构建可实现指数级复合增长的智能体平台。 ## TL;DR 演讲系统性拆解了吉利德 GMIND 平台的架构哲学，倡导让大模型负责“广度推理（Reason Broadly）”理解意图，而将具体执行固化在“确定性 Temporal 持久化工作流（Act Precisely）”中。平台通过类型契约、上下文包与持久化引用实现多模态能力的深度解耦，并提供对话（Chat）与画布（Canvas）双交互路径以满足不同业务习惯。实战案例表明，该模式在合规文档生成与研究模拟中带来了显著的效能跃升（文档生成时间从数小时缩短至 5 分钟），同时配合明确的平台治理与共享资金模型，推动企业 AI 平台实现实用价值超越系统复杂度的可持续演进。 ## Atmosphere & style 整场演讲流露出严谨、务实且极具工程洞察力的专业基调。主讲人毫不避讳团队早期追逐“智能体炒作”所遭遇的效率瓶颈，通过翔实的实测对比和架构图解，冷静剖析了纯 Agent 模式与确定性工作流的权衡边界。叙事逻辑清晰连贯，既有微观层面的代码解耦与技术栈迁移实操，又有宏观层面的企业治理与负责任 AI 思考，给观众带来一种脚踏实地、拨开行业炒作迷雾的厚重获得感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) GMIND 的核心架构理念是追求能力的复合效应（Compounding），新接入的专业能力通过智能体与现有模块有机组合，驱动平台整体价值指数级增长，而非简单的功能线性堆叠。 - [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) 在企业应用中，“多模态”被拓展为涵盖自然语言、文档、图像及半结构化数据库的业务证据，通过上下文包（Context Packets）、类型化 JSON 和持久化引用实现精准路由，避免粗暴拼接巨型 Prompt。 - [03:08](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=188s) 前端提供 Chat（自然语言探索）与 Canvas（可预测画布控件）双路径交互，同时在后端统一连通至鉴权、遥测及 Temporal 执行底座，兼顾灵活性与企业用户的确定性诉求。 - [03:08](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=188s) 所有功能均封装为能力契约（Capability Contracts），各业务团队只需专注于领域逻辑，彻底消除点对点网状互联造成的耦合风险。 - [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) 严格遵循整洁架构（Clean Architecture），将大模型与编排引擎外置于外围设施，使得团队仅耗时一天便无缝完成了从 LangGraph 到 Vercel AI SDK 与 Temporal 编排层的迁移。 - [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) 确立“广度推理，精确行动（Reason broadly, act precisely）”的执行边界：让大模型负责意图理解与宽泛推理，而将复杂的关键业务动作固化在具备容错持久性的 Temporal 工作流中。 - [09:43](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=583s) 在研究模拟（Study Simulation）场景中，以定向工作流替代智能体拼装庞大复杂 JSON 的试错死循环，将数据准备耗时从数分钟急剧缩短至 10 秒以内。 - [09:43](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=583s) 合规文档生成案例表明，盲目使用纯智能体导致耗时长且充满幻觉，改用结构化确定性工作流后，生成时间由数小时降至 5 分钟，效率提升幅度跃升至 75%。 - [13:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=814s) 标准化契约与持久化引用让跨业务线联动无缝发生，主管智能体可在无专门 UI 的情况下将半结构化 Excel 文件自动扇出（Fan-out）为多个并行模拟工作流。 - [15:58](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=958s) 从单一指标优化迈向多维决策平衡，在临床试验规划中统筹考量速度、成本、可行性约束与竞品管线，兼顾统计学严谨模拟与宏观商业推理。 - [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) 组织变革上明确划分“平台所有者”（负责通用契约、底座基础设施与安全）和“能力所有者”（深耕具体业务逻辑），并配套设立跨领域共享价值的平台级资金预算模型。 - [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) 构建企业级多模态 AI 平台的终极准则是：以专业化、确定性与耐用性建立信任底座，依托稳定技术契约实现解耦，使实用价值的增长速度大幅超越系统复杂度的攀升。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) GMIND 平台概述与多模态架构设计 在演讲的开头，来自吉利德（Gilead Sciences）临床开发部门 AI 研究中心的总监 John Austin 介绍了他们团队正在研发的**多模态药物开发洞察网络（Gilead’s Multimodal Insight Network for Drug Development，简称 GMIND）**。GMIND 是一个吉利德内部共享的 AI 应用平台，其核心架构追求的是**复合效应（Compounding）**：每当有新的专业能力（Capability）接入平台时，它不只是简单地增加一个独立功能，而是通过智能体（Agent）的粘合作用，与已有能力产生有机结合，从而让整个平台的综合能力呈现指数级增长。 ```mermaid flowchart TD accTitle: GMIND 多模态数据整合与路由架构 subgraph Data ["多模态证据输入"] NL["自然语言与记忆"] DOC["文档与 PDF"] IMG["图片/图表/表格"] DB["半结构化数据库"] end subgraph Alignment ["证据对齐与架构契约"] CP["Context Packets<br/>上下文包"] TJ["Typed JSON<br/>类型化 JSON"] DR["Durable References<br/>持久化引用"] end subgraph Execution ["工作流路由与执行"] RTR["Fit-for-Purpose Routing<br/>适配路由"] WF["Specialized Workflows<br/>专业工作流与工具"] end Data --> Alignment Alignment --> RTR RTR --> WF ``` John 澄清了制药行业对“多模态”的传统误解——很多人以为这仅仅是将病理图像、医学影像或组学数据（OMIX）融合进同一个大模型中。而在 GMIND 的应用体系里，“多模态证据”涵盖了自然语言请求、历史记忆、PDF 文档、图像、表格、半结构化数据库记录以及其他能力的输出。为了将这些多模态数据转化为精准的业务行动，GMIND 采用了严谨的对齐与传输机制： - **Context Packets（上下文包）**：附带数据来源（Provenance），提供紧凑且可解释的上下文； - **Typed JSON（类型化 JSON）**：当工具需要精确接口规范时使用； - **Durable References（持久化引用）**：对于尺寸过大、无法直接放入 Prompt 或在工作流中传递的产物，使用持久化引用进行追踪； - **Fit-for-Purpose Routing（适配路由）**：根据具体任务将数据分发给最合适的工具或工作流，杜绝“将所有数据拼接进一个巨型 Prompt”的粗暴做法。 --- ### [03:08](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=188s) 交互双路径与能力契约（Capability Contracts） GMIND 在前端为用户提供了统一且连贯的交互空间，并设计了**双途径交互模式**： 1. **Chat 路径（对话模式）**：用户可以通过自然语言描述目标、调整需求并附加证据，由系统自动理解意图并调用经过授权的工作流； 2. **Canvas 路径（画布模式）**：针对高度可重复的标准化业务流程，用户可以直接在迷你画布界面上使用专用的控制控件触发对应工作流（如文档编辑等）。 ```mermaid flowchart LR accTitle: GMIND 双途径交互与后端统一架构 subgraph Interactive ["前端交互路径"] Chat["Chat 路径<br/>自然语言探索/意图理解"] Canvas["Canvas 路径<br/>可视化控件/可预测流程"] end subgraph Contract ["能力契约层"] CC["Capability Contract<br/>类型受控契约"] end subgraph Shared ["统一后端基础设施"] Auth["统一授权与鉴权"] Tel["遥测与监控 (Telemetry)"] Dur["Temporal 持久化执行"] end Chat --> CC Canvas --> CC CC --> Shared ``` 这种双路径设计的背后，是吉利德对制药行业用户心理的深刻洞察：虽然一部分用户喜欢使用自然语言进行探索性交互，但仍有大量用户对纯智能体交互持怀疑态度，他们更青睐带有极少“AI 痕迹”、每一步都高度可预测的定制化软件。GMIND 将这两条路径连通至完全相同的后端基础设施（统一的授权、遥测及 Temporal 持久化执行能力），避免了平台碎片的困境。 无论是工具、API 还是智能体，所有部署到 GMIND 的功能都被封装为**能力契约（Capability Contract）**。新功能的开发团队只需专注于自身业务领域的专属逻辑，无需重复构建底层的通用基础设施；同时，能力之间通过类型契约、上下文包和持久化引用进行通信，消除了点对点网状互连（Point-to-point mesh integration）的耦合风险。 --- ### [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) 整洁架构与 Temporal 持久化工作流 为了防止系统在组合过程中走向高度耦合，GMIND 严格遵循了 Uncle Bob（Robert C. Martin）提出的**整洁架构（Clean Architecture）**原则。核心业务逻辑和领域模型被妥善保护在稳定的 Ports、DTOs（数据传输对象）和工作流契约之后，而诸如大模型、记忆库、编排引擎和数据库等组件则全部被推到了最外层的“外围（Periphery）”。依赖关系严格向内单向指向。 > "That separation let us move from LangChain/LangGraph to Vercel's AI SDK and Temporal orchestration without redesigning any capabilities in only a single day's worth of work." > > （这种架构隔离让我们在仅一天的时间内，就完成了从 LangGraph 迁移至 Vercel AI SDK 和 Temporal 编排引擎的改造，且没有重构任何现有的业务能力。） ```mermaid flowchart TD accTitle: GMIND 整洁架构分层图 subgraph Outer ["外围设施 (Periphery)"] LLM["大语言模型 / AI SDK"] DB["数据库 & 记忆库"] Orch["Temporal 编排引擎"] end subgraph Core ["核心业务层 (Core)"] Ports["稳定接口 (Ports) / DTOs"] BizLogic["领域逻辑 & 能力契约"] end Outer -->|依赖单向指向内层| Core ``` 在 GMIND 中，一个**可调用的工作单元**不是普通的函数或 API，也不是简单的 MCP 调用，而是一个**Temporal 持久化工作流（Durable Workflow）**： - **执行跨度极大**：工作流的运行时间可以从零点几秒、数小时延伸至数年，具备极高的容错性与持久性； - **确定性流程降本增效**：将复杂的业务过程收敛为显式步骤，消除了纯智能体反复规划、重试所带来的高延迟与不稳定； - **优化路径清晰**：当根据反馈需要改进系统时，团队可以精准调整确定性的步骤逻辑，而无需漫无目的地微调 Prompt 或 Skills。 系统由此实现了清晰的分工：**让 LLM 负责广度推理（Reason Broadly），理解意图与上下文；让适配的 Temporal 工作流负责精确执行（Act Precisely），掌控复杂业务动作。** --- ### [09:43](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=583s) 实践案例：研究模拟与文档生成工作流重构 John 共享了两个极具说服力的实际案例，展示了从“纯 Agent 模式”转向“定向 Temporal 工作流”所带来的巨大性能提升： #### 案例一：研究模拟（Study Simulation） 在药物开发中，研究模拟的输入参数往往是一个数千行的复杂 JSON 对象。 - **纯 Agent 模式**：模型需要在对话上下文和数据库查询之间陷入循环，自己去提取、解析并拼装这个庞大的 Payload。虽然校验机制能判断 JSON 结构是否合法，但无法保证输入数据的完整性和业务准确性。智能体频繁陷入自我修复的死循环中，准备时间长达数分钟且结果不可靠； - **定向工作流模式**：工作流直接提取权威数据源，按 Schema 进行精准对齐，应用用户自定义的参数更新后持久化输出。准备时间从**几分钟大幅缩短至 10 秒以内**，数据质量显著提升。 #### 案例二：合规文档生成（Document Generation） 制药行业的文档通常基于版本受控的严格模板。 - **纯 Agent 模式**：团队最初上了“智能体热潮（Agentic Hype Train）”，给智能体赋予广泛的工具和数据接口去自由生成文档。结果智能体耗费数小时不停规划与运转，烧掉大量 Token，生成的章节却充满了幻觉和数据缺失。人工核对与纠错的成本几乎抵消了所有收益，时间节省率**甚至达不到 25%**； - **定向工作流模式**：改用按模板拆分章节的结构化工作流，利用智能体并行处理各章节内容、审查质量与准确性后进行拼装。生成时间从**数小时缩短至 5 分钟**，时间节省率直接攀升至 **75%**。 > "The distinction is not AI versus no AI. It's where you allow the freedom." > > （核心区别不在于‘用不用 AI’，而在于‘你在何处给予自由’。判断、推理和内容生成交给智能体，而对于几十年积累下来的成熟业务流程，必须用确定性的工作流去约束。） --- ### [13:34](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=814s) 能力的组合复合与跨业务线应用 当每一个专业能力都通过标准契约暴露出来后，GMIND 实现了真正的“1 + 1 > 2”的复合增长效果： ```mermaid flowchart TD accTitle: GMIND 复合能力组合示例 subgraph Inputs ["多源数据输入"] MonData["现有临床研究监测数据"] Excel["半结构化 Excel 文件"] end subgraph CoreEnv ["GMIND 契约与编排环境"] SupAgent["Supervisor Agent<br/>意图解析与分发"] SimWF["研究模拟工作流 (Study Sim)"] end subgraph Artifacts ["持久化产物与迭代"] DurRef["持久化文档引用<br/>(Durable References)"] Revise["后续多轮修改与完善"] end MonData --> SimWF Excel --> SupAgent SupAgent -->|自动扇出/批处理| SimWF SimWF --> DurRef DurRef --> Revise ``` - **数据无缝联动**：临床研究监测数据在无需任何硬编码点对点集成的情况下，直接被智能体提取并注入到新研究的模拟工作流中； - **持久化引用支持持续迭代**：文档生成不再是一次性的“初稿生成”，持久化存储引用让后续的工作流能够重新检索、继续补充和修改文档； - **无 UI 的批处理能力扩展**：当用户上传包含多行模拟参数的半结构化 Excel 文件时，主管智能体（Supervisor Agent）会自动解析数据并将其扇出（Fan-out）为多个并行运行的工作流，在没有专门编写批处理 UI 的情况下实现了批量模拟。 --- ### [15:58](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=958s) 多维优化与组织管理变更 在谈及平台未来的演进方向时，John 强调 GMIND 正从单一维度的优化转向多维度的商业决策支持。以临床试验的操作规划（如站点选择、区域部署、激活顺序）为例： - 纯粹追求**速度**会盲目增加站点； - 加入**成本**考量后，最快方案不再是首选； - 引入**可行性约束**（如单个城市的患者饱和度限制）、**流行病学数据**、**竞品时间线**以及**公司管线优先级**后，整个可行性边界（Feasibility Frontier）会发生深刻改变。 GMIND 利用严谨的模拟工具保证统计学上的绝对精准，同时利用智能体去理解和权衡复杂的业务因素。 在演讲最后，John 分享了企业推进这类平台时在**组织与管理（Change Management）**上的经验：传统企业习惯于“一个产品、一个领域、一个负责人”的模式，而复合型 AI 平台打破了这一边界。吉利德成功实施变更的关键在于明确权责分工——**平台所有者（Platform Owner）**负责共享契约、运行体验、架构标准与安全合规；而**能力所有者（Capability Owners）**则分布在各个业务部门，专注于各自领域的逻辑开发。双方互不干扰，共同驱动平台能力的快速迭代。 - **核心原则总结**： Reason broadly（广度推理）, act precisely（精确行动）, and compound through contracts（通过契约实现复合增长）。 - **技术栈迁移**：成功在 1 天内将编排层由 LangGraph 迁移至 Vercel AI SDK + Temporal，验证了整洁架构的优越性。 - **性能实测**：合规文档生成时间由数小时降至 5 分钟，效率提升从低于 25% 跃升至 75%；研究模拟数据准备时间从几分钟缩短至 10 秒。 <ElicitationsGroup message="如果您想进一步了解 GMIND 的架构细节，可以探索以下方向："> <Elicitation label="深入了解 Temporal 在 AI 工作流编排中的作用" query="请详细解释 Temporal 在 AI 智能体和工作流编排中扮演的角色，以及它相比纯 Prompt 链有哪些优势？"/> <Elicitation label="对比整洁架构在 AI 平台开发中的设计模式" query="请对比说明整洁架构（Clean Architecture）在现代 AI Agent 平台架构设计中的具体应用与落地模式。"/> </ElicitationsGroup> ### [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) 领域自治与平台治理：监管约束下的协同与资金模型 在深入探讨系统落地时，业务领域与平台层级之间的职责边界至关重要。处于业务一线的团队与具体的业务领域高度绑定，他们直接对业务意图、输入数据、质量评估标准、人工复核机制以及该领域的负责任使用承担最终责任。正因如此，这些业务专家的视角往往高度聚焦，通常并不会过多关注其他业务领域的动态与诉求。然而，一个整体性的多模态智能平台绝不能陷入各自为政的孤岛，必须有人站在全局视角，审视并把控跨业务领域的整体架构与连接方式。 能力间的组合调用并非无序发生，它必须置于严格的平台治理与政策约束之下。明确哪些智能体之间的移交（handoffs）是被核准的、哪些预期的下游用途是被允许的，构成了平台安全的核心。在生物医药这种受到严密监管的高门槛行业以及 Gilead 内部，对负责任 AI（Responsible AI）的践行有着近乎严苛的要求，系统必须部署充足的安全护栏、控制措施与标准流程，以确保平台始终“创造价值而非带来危害”（"deliver value and not harm"）。 ```mermaid flowchart TD accTitle: 平台治理与业务领域协同架构 subgraph Governance["平台级治理与支撑"] P["平台政策与安全护栏<br/>(核准移交与合规用途)"] F["共享价值资金模型<br/>(平台级统筹投入)"] end subgraph DomainA["业务领域 A"] DA["领域数据与评估"] RA["人工复核 / 负责任使用"] end subgraph DomainB["业务领域 B"] DB["领域数据与评估"] RB["人工复核 / 负责任使用"] end P --> DomainA P --> DomainB F -.-> P ``` 这种平台级定位直接冲击了传统企业的财务与立项逻辑。支撑模型与资金分配机制必须如实反映这种“共享价值”，绝不能再沿袭为单个孤立能力分别批复预算的旧模式，因为平台绝大部分的工程重击与沉淀工作都发生在底座层级。对于那些在传统思维根深蒂固的大型成熟企业中探索落地的团队，在项目初期尽早与管理层就这一资金与协同模型达成共识，是决定平台生死的关键一步。 - 领域责任范畴：业务领域团队专注于自身闭环，严格负责数据源质控、评估框架、专家介入的人工复核（Human Review）及本领域的合规使用。 - 组合治理机制：系统严格管制智能体间的交接路径（Approved Handoffs）与获批的使用场景（Intended Uses），杜绝不受控的级联调用。 - 资金模式转变：破除传统按单个功能/工具切分预算的立项方式，转向由平台层级统一承接资金与技术资产沉淀的共享投资模型。 --- ### [19:07](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1147s) 可组合智能的终极范式：让实用价值超越系统复杂度 在整个架构哲学的总结阶段，核心逻辑得到了极致提炼：构建企业级多模态 AI 平台的终极目标，绝非盲目堆砌更多的智能体（Agents）、工具（Tools）、功能特性（Features）或海量数据。如果缺乏坚实的系统信任度与运行可靠性，盲目增加组件数量只会带来系统复杂度的指数级失控，并制造源源不断的故障风险。唯有专业化（Specialization）、确定性（Determinism）与系统耐用性（Durability），才能真正沉淀出底层信任与可靠质量；而在此基础之上的灵活组合，才能释放出乘数级的增量价值。 能否同时兼顾高可靠性与可组合性，完全取决于企业的技术架构与组织系统设计。架构层面的最佳实践应当是清晰分工：让 AI 充分发挥其广谱推理（Broad Reasoning）能力，去理解全域证据与用户意图，并在全局中检索匹配最合适的能力组件；但对于任何复杂、精密、要求严谨的具体执行动作，必须将其严格封装在稳定耐用、因地制宜的确定性工作流之中。各个能力模块之间则通过极其稳固的技术契约（Stable Contracts）进行解耦与通信，从而免除工程师团队在功能扩展时反复修补、救火与重构的无休止内耗。 这套严密的架构准则，正是大型多模态智能平台实现“实用价值增长快于系统复杂度攀升”的根本方法论。随着现场计时器归零，演讲在热烈的学术与工程探讨氛围中画上句号，未尽的技术细节与互动探讨则顺畅移步至会下展开。 - 信任与价值的来源：专业化、确定性与耐用性构筑系统的可信底座，跨模块的安全组合则催生几何级数的业务价值。 - 动静分离的执行边界：将宽泛的推理决策（理解意图、选择工具）交由 AI，将严密的专业操作固化在定制耐用的确定性工作流中。 - 契约驱动解耦：通过稳定契约定义模块间接口，彻底斩断能力扩张时引发的代码回溯与维护连锁反应。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=4s) 每当有新的专业能力接入平台时，它不只是简单地增加一个独立功能，而是通过智能体的粘合作用，与已有能力产生有机结合，从而让整个平台的综合能力呈现指数级增长。 — John Austin > [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) That separation let us move from LangChain/LangGraph to Vercel's AI SDK and Temporal orchestration without redesigning any capabilities in only a single day's worth of work. — John Austin > [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) 这种架构隔离让我们在仅一天的时间内，就完成了从 LangGraph 迁移至 Vercel AI SDK 和 Temporal 编排引擎的改造，且没有重构任何现有的业务能力。 — John Austin > [06:15](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=375s) 让 LLM 负责广度推理（Reason Broadly），理解意图与上下文；让适配的 Temporal 工作流负责精确执行（Act Precisely），掌控复杂业务动作。 — John Austin > [09:43](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=583s) The distinction is not AI versus no AI. It's where you allow the freedom. — John Austin > [09:43](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=583s) 核心区别不在于‘用不用 AI’，而在于‘你在何处给予自由’。判断、推理和内容生成交给智能体，而对于几十年积累下来的成熟业务流程，必须用确定性的工作流去约束。 — John Austin > [15:58](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=958s) Reason broadly, act precisely, and compound through contracts — John Austin > [18:02](https://www.youtube.com/watch?v=TvxYZhG1iTk&t=1082s) deliver value and not harm — John Austin ## Mentioned John Austin, 吉利德 (Gilead Sciences), GMIND (Gilead’s Multimodal Insight Network for Drug Development), Ai4, Robert C. Martin (Uncle Bob), Temporal, LangChain, LangGraph, Vercel AI SDK, Clean Architecture (整洁架构) ## Video description John Austin - Director of Translational AI, Gilead Multimodal AI systems often try to scale by adding ever more data sources, agents, and tools. But a larger bundle of components does not automatically produce more useful or reliable capability. The greater opportunity is compounding: each new specialized capability should not merely add its own value, but expand what existing capabilities can accomplish through safe combination. The architectural aim is not linear accumulation, but exponential growth in what the platform can do as those combinations multiply. This talk presents the thinking behind GMIND—Gilead’s Multimodal Insight Network for Drug Development—a agent-first platform designed to grow from individual, high-value use cases into an expanding system whose capabilities become more powerful together. Drawing on lessons from document generation, study simulation, and other drug-development workflows, it explores how GMIND balances deep specialization with cross-capability composition, agentic flexibility with prescriptive process, and rapid expansion with the reliability and control required in pharmaceutical R&D. Attendees will leave with practical principles for building multimodal AI platforms whose capabilities grow faster than their component count—not merely systems that accumulate more disconnected features. [Source](https://www.youtube.com/watch?v=TvxYZhG1iTk)

Source quotes

[0:04] John Austin: 每当有新的专业能力接入平台时，它不只是简单地增加一个独立功能，而是通过智能体的粘合作用，与已有能力产生有机结合，从而让整个平台的综合能力呈现指数级增长。 [6:15] John Austin: That separation let us move from LangChain/LangGraph to Vercel's AI SDK and Temporal orchestration without redesigning any capabilities in only a single day's worth of work. [6:15] John Austin: 这种架构隔离让我们在仅一天的时间内，就完成了从 LangGraph 迁移至 Vercel AI SDK 和 Temporal 编排引擎的改造，且没有重构任何现有的业务能力。 [6:15] John Austin: 让 LLM 负责广度推理（Reason Broadly），理解意图与上下文；让适配的 Temporal 工作流负责精确执行（Act Precisely），掌控复杂业务动作。 [9:43] John Austin: The distinction is not AI versus no AI. It's where you allow the freedom. [9:43] John Austin: 核心区别不在于‘用不用 AI’，而在于‘你在何处给予自由’。判断、推理和内容生成交给智能体，而对于几十年积累下来的成熟业务流程，必须用确定性的工作流去约束。 [15:58] John Austin: Reason broadly, act precisely, and compound through contracts [18:02] John Austin: deliver value and not harm

Source quotes

[0:04] John Austin: Reason broadly, act precisely, and compound through contracts. [0:04] John Austin: app store [1:34] John Austin: multimodal AI [4:39] John Austin: capability contract [9:50] John Austin: Use the LLM to understand your input and your intent and to select capabilities, but don't make it the transport and transformation layer for your data. [11:27] John Austin: AI versus no AI. [18:02] John Austin: responsible AI use [18:02] John Austin: deliver value and not harm. [19:07] John Austin: more agents, tools, features or data. [19:07] John Austin: constantly go back and fix and update and improve [19:07] John Austin: become more useful than they become complex

Original transcript

[0:04] Good morning. I'm as you said, John from Gilead  Sciences. I'm a director in the AI research [0:09] center within clinical development there. And  the work I'm sharing today reflects some of [0:13] my own contributions, but more importantly,  the contributions of many of my colleagues [0:17] across development and IT at Gilead. I'm here to  share elements of the architecture and lessons [0:23] learned uh from the ongoing development of  Gilead's multimodal insights network for drug [0:28] development which we prefer to call Gmind for  short. Gind is an internal AI application. Um [0:36] it for it's it's it's like a it's a basically like  a shared application platform which I'll go into. [0:43] Each capability deployed to the platform  solves a real problem on its own. But the [0:46] architectural goal of Gmind is compounding. So  where domains intersect because we have a lot [0:52] of different departments within development, a  lot of different goals within drug development. [0:57] Each new capability should expand what the  rest of the system can accomplish, not just [1:01] merely add another feature. So we're not trying  to make like an app store where we have separate [1:05] apps. the the the the real benefit that we're  trying to achieve is through agents being able to [1:10] glue and then be able to have um increasing  capability beyond even the original expectations. [1:17] I'll share the design choices and  lessons learned behind that approach [1:21] including why greater agent autonomy very often  made outcomes slower and less reliable and the [1:27] principles we now use which is reason broadly, act  precisely and compound through contracts. So first [1:34] uh being in this multimodal track uh let me  define what multimodal means in this architecture. [1:40] So in pharma multimodal AI usually evokes kind of  images of pathology images or other medical images [1:47] or OMIX data or other things like that being fused  together inside a model. But Gmind is using a [1:53] wider sort of application system definition. It's  not just about running a single kind of model. [1:58] So multimodal evidence in Gmind um means natural  language requests, memories, documents, PDFs, [2:05] images, tables, figures, uh database,  semi-structured database records, um [2:11] and outputs from other capabilities. The important  step is aligning the evidence for action. [2:16] So we use context packets with uh prominence  for compact interpretable context and typed [2:22] JSON when a tool requires an exact interface.  Durable references when an artifact is too large [2:29] or it needs to be very precise and you can't  persist it through prompts and workflows. [2:35] And we also use fitforpurpose routing for  the right tool or workflow. So multimodal [2:40] evidence is a starting point. It's not really our  prompt strategy. So we do not append everything [2:45] into one enormous prompt and ask an LLM or team of  agents to sort things out. So what we do instead [2:50] is we develop reusable workflows and specialized  tools that meet specific business needs. The AI [2:56] model reasons across the system. Yet what we do  is we use specialized capabilities and workflows [3:02] tailored to specific business processes to provide  dependable actions and reusable sets of work. [3:08] So, Gine provides a unified cohesive  workspace with two main interaction paths. [3:14] This interface is not very unique to us. This  is probably, you know, chat next to some kind of [3:19] canvas. Um, but this is what we do. So, there's  two main interaction paths. Through the chat, [3:24] a user can describe an objective, refine, attach  evidence. The system interprets the intent, [3:28] gathers context, and runs the approved  workflows or tools to get them to their goal. [3:34] And then on the right, you're going to see the one  of the canvases. So in one of the mini canvases, [3:38] users can invoke workflows directly through  controls designed for repeatable business process. [3:43] So this is an illustrative example  of document editing. I unfortunately [3:47] couldn't get through our legal uh one of  our more unique workflows though like this. [3:52] Each workflow that's in this kind of canvas  side is tuned to a very specific type of work. [3:57] So GMAN does not only meet users needs through  purely agentic interactions. Some uh users do [4:04] want that exploration and being able to use  natural language to interact with uh Gmind. [4:09] However, others and a large number of others  still are hesitant and concerned and they want [4:15] fitfor-purpose software with minimal visible AI  and very predictable steps even though behind the [4:20] scenes there is often a lot of AI being employed.  So both paths reach the same um authorization [4:28] artifacts, telemetry, durable execution  capabilities. We don't build separate backends. [4:34] This gives people a much smoother path into using  and trusting AI without fragmenting the platform. [4:39] So the common unit behind both of  these paths is the capability contract. [4:45] In Gmind a capability is bounded domain behavior  exposed through a type governed contract. [4:51] It may be a tool, durable workflow, an  external API or or even a specialized agent. [4:58] a new capability implements that contract rather  than constructing a new application around itself. [5:04] Through that interface, it inherits  conversation and canvas access, uh, [5:09] authorization, intended use, temporal  execution, memory, and just lots of other [5:14] foundational things to the platform. So, what  this does is it really changes the economics [5:18] of how we can build and deliver the application  internally. Each delivery team can focus on their [5:23] specialized business domain logic and they don't  need to worry about any of that foundational stuff [5:30] and also avoids point-to-point mesh integration of  capabilities. The capabilities communicate through [5:35] those type contracts context packets and durable  artifact references and these callable workflows. [5:41] So parent like parent workflows can achieve can  sorry retrieve uh evidence prepare the validated [5:48] inputs run simulations and pass the results on  and it can do all of this within those steps [5:54] and pass it on to other workflows without ever  having needed to know that those other ones exist. [6:01] So basically what we're doing is we're making  lots of very specific tools and then having the AI [6:05] merge them all together. This is where composition  begins. It's add a specialized capability once [6:09] and it gains multiple callers and multiple  possible collaborators that grow over time. [6:15] So clean architecture as published by uncle Bob  is hopefully very familiar to anyone here who's [6:21] in software engineering. Its value here is that it  prevents composition from turning into coupling. [6:27] So what we we strictly adhere to clean  architecture principles and the development [6:31] of Gmind and it allows us to grow with much less  friction. So the core so for instance the core use [6:37] case and domain logic sit behind stable ports  data transfer objects and workflow contracts [6:44] things like the models memory orchestration  databases a lot of things that [6:48] are often considered really core to the to to the  application those are all in the periphery they [6:53] they they don't really matter what matters is  the business logic and that that business logic [6:57] is consistent so the dependencies point inward  and code that changes together stays together [7:03] so that separation let us move from as an example  example, moving from Langraph to Verscell's AI SDK [7:08] in temporal orchestration without redesigning any  capabilities in only a single day's worth of work. [7:15] It also lets teams across departments develop  capabilities in parallel while respecting all [7:20] of these same boundaries and not requiring  them to stay in contact. So, we're constantly [7:24] releasing new capabilities across departments and  we don't need to worry about them interfering or [7:29] degrading uh one of the other uh capabilities. So,  the concentric rings are not the novel part here. [7:34] The consequence of using clean architecture is  what's novel at least for systems with at Gilead. [7:40] The platform evolves without forcing every  previously integrated AI capability to [7:44] evolve with it. That is what is enabling  basically near plug-and-play integration [7:49] and allows our small development teams across  multiple departments to work together on this. [7:57] These stable contracts become especially powerful  when the callable capability is a proven process. [8:03] So in GM mind we treat a temporal workflow  as a tool. For those of you not familiar with [8:09] temporal for orchestration, I highly recommend  familiarizing yourself with it. This is really [8:13] completely transformed how Gmind is able to carry  out work on a user's behalf. So the callable unit [8:19] of work in Gind isn't a function or API call. It's  not a direct MCP call. It's a durable workflow. [8:26] These workflows can span fractions of a  second, hours. They can go all the way up [8:30] to years if necessary and be extremely durable.  So an agent, canvas or parent workflow calls [8:38] it the same type contract when executing the  same task. In an example, inside that boundary, a [8:44] workflow may validate intent, access, and inputs,  execute some required steps, verify the quality, [8:51] and then persist that durable reference to pass  it on. And that creates the three benefits uh [8:56] on the left. Explicit steps are hand and handoffs  reduce variability. Known paths they avoid [9:05] repeated planning loops, retries and latency.  And when feedback identifies an improvement, [9:10] it's most often that what we're doing  is we're refining these workflows and [9:13] these deterministic steps rather than  trying to improve the prompts for skills. [9:18] So the division is deliberate. The AI is  reasoning about uh what should happen. [9:23] But fitfor-purpose workflows are controlling  how we carry out those complex actions. [9:29] We ensure the reusabil the reusable capabilities  and handoff contracts. So it's really important [9:34] the contracts and between all the different  capabilities that's what we're really focused [9:38] on keeping very rigid and stable and we don't  have to worry about the prompts uh keeping them [9:43] stable. So the study simulation is another  capability that really exemplifies this. So [9:50] one of the main areas where AI can help uh improve  drug development is improving study simulations [9:57] for forecasting and decision-m. A study simulation  input is is basically a precisely defined JSON [10:04] object that can be thousands of lines long in the  prompt driven path where we're not trying to just [10:09] drive this through a much more directed workflow.  Then we were doing that before we moved to these [10:14] directed workflows. The model had to retrieve,  interpret and construct that large payload from [10:19] conversational context in database access in a  loop. So validation could tell us in our loop that [10:26] the object was structurally valid, but not always  that the intended input was complete or correct [10:33] also. And then when the structural valid  verification fails, the agent would loop [10:38] while trying to repair it. That preparation  often took several minutes and left a lot of [10:42] uncertainty about quality and repeatability. The  workflow version prepares the input directly, [10:48] retrieves authoritative data, aligns it to a  schema. It applies any user specific updates [10:53] to the data for the simulation that you want  to run and then it will eventually persist it [10:59] and pass it on into the simulation's input. That  reduced preparation time from many minutes down [11:05] to 10 seconds and also really inc imp improve the  the quality. So the lesson here is pretty simple. [11:11] It's use the LLM to understand your input and  your intent and to select capabilities, but [11:17] don't make it the transport and transformation  layer for your data, right? Unless it's [11:21] really the very last resort. So, in document  generation, uh the goal is to generate uh [11:31] document drafts using generative AI to aid in that  especially in that first draft generation process. [11:36] These documents especially in pharma are often  generated from version controlled templates. [11:42] So we first tried a more permissive  pattern. You know probably a lot of [11:46] you have tried this at first jumping on the like  the agentic hype train agents can do everything. [11:51] We broad capabilities tools and data and asked  them to produce the document. It took many hours. [11:57] It churned and planned. It burned through lots  of tokens way too many tokens and produced [12:02] sections with hallucinations missing data. um  basically the human verification burden of reading [12:09] it and figuring out all the places it messed up  that erased all the benefit almost. So in one [12:14] document type that uh when we evaluated and worked  with business users it didn't even meet it didn't [12:21] even reach a 25% time savings threshold with all  the validation and correction that had to be done. [12:27] So we couldn't proceed there but then we went to a  much more directed workflow. it still uses agents [12:32] very heavily um but more and more of a prescribed  process. So using the template, having predefined [12:37] ways to deal with documents, understanding how the  data usually flows into these different sections, [12:44] we were able to, you know, break it down into  sections, do a lot of work in parallel, um, [12:50] review them, review the quality, and accuracy,  revise it, and assemble the document. [12:54] And now more deterministic workflow  that drastically increased the quality, [12:58] and it only ran for it only needs to run for 5  minutes instead of for a couple hours. And in that [13:02] same document type in it reached 75% time savings  versus not even 25%. So bless you. The uh the [13:11] distinction is not AI versus no AI. It's where you  allow the freedom. Right? So judgment, reasoning, [13:17] content generation, all of that remains agentic.  But where we have business processes that you know [13:23] we've been making these documents for years, for  dozens of years where we have defined business [13:28] process and we know how to do things well, it's  much better to just build out that process. [13:34] So this is kind of moving on. This is what that  composition and nonlinear growth and overall [13:39] capability looks like in practice. So I said that  we what we were building these specific workflows [13:43] and there were lessons learned from them but  we also wanted when we build more workflows for [13:46] them to start working together and you know  1 plus 1 should be more than two basically. [13:51] So in the first example um in the top left right  [sighs] data connected for monitoring current [13:58] studies can also be selected and used by the agent  and then provided uh as part of the input for new [14:04] study simulation. Right? Right. So then that  monitoring investment enriched the simula the [14:10] the enriched planning with in collaboration with  the stud the simulation uh work without having to [14:16] make any tailor made pointto-oint integration or  considering ahead of time. Additionally internal [14:21] document and artifact storage make outputs uh  addressable. So durable references let later work [14:29] retrieve continue and revise the document enabling  the capability to go not go beyond just first [14:35] draft generation to like refining and improving  that document over and over again. So persistence [14:40] at the data storage layer became a reusable  capability uh rather than just the end of first [14:45] draft generation. So furthermore, file uploads  they create another path into simulation. So [14:58] at its base, the simulation just wants  to run one simulation. However, if you [15:02] put like uh send upload something like a  semi-structured Excel with maybe rows of [15:08] different simulations you want to run that the  the agent the supervisor agent can take that [15:13] figure out how to turn that into multiple runs fan  that out to several workflows and then now without [15:19] having to pre-plan it now you've created this  batch processing and you didn't have to build any [15:22] specific UI you didn't have to specifically make  a um an input for it so none of the combinations [15:29] here are magic right And that's the point. So  governed interfaces make the data artifacts and [15:35] workflows available for approved reuse. The model  reasons about when a resource is useful and usable [15:41] and then the fitforpurpose capability acts.  So each addition solves its own problems, [15:46] but it also gives the existing system another  resource to combine. So the effect becomes more [15:51] consequential as those reusable capabilities  begin to shape a shared optimization problem. [15:58] I'm probably going to have to go a little fast  here, but um okay, so G Mind is still in very [16:02] active development with a large road map. I  wanted to share one of the things what we're [16:06] going towards that really shows how combining all  these multimodal inputs brings value and high ROI. [16:11] So one of the things we're really focusing on is  enriching that simulation and the ability to drive [16:17] business value and decision-m even further. So  in a naive sense, if we just optimize for speed, [16:23] so an operational plan is like which sites  do you select, which people do you work with, [16:27] in what regions do you deploy them, in what  order do you activate sites? So it's like the [16:30] plan of carrying out a study. If we add a bunch  of sites, the model's going to say, "Oh, great. [16:35] You're going to enroll faster." But of course, if  you add cost, then you're going to see that maybe [16:40] the fastest scenario isn't what you prefer, and  there's a clear trade-off between speed and cost. [16:44] You add feasibility. So for instance, there are  limits to the number of patients in a region. You [16:49] can't just keep putting more and more sites  into Los Angeles. There's there's definitely [16:52] amount of saturation you get to and there's  many other operational realities to consider [16:58] and a lot of those will start removing attractive  scenarios and change that feasibility frontier. [17:02] You can add epidemiology, competitive timing,  portfolio priorities and they all change things. [17:09] So it's not just about feeding more data into one  optimization problem. the data and how it's used [17:14] and the business considerations really change and  it really is helpful to have that agent trying to [17:18] interpret how to use the data even though a lot  of the things like the simulation itself is very [17:23] uh very locked because we we want it  to be very statistically rigorous. [17:30] So going to move on to this is sort of the  non-technical part. It's where the change [17:34] management side of things. Traditional  enterprises often assume one product, [17:39] one domain, one owner, one support path.  A compositional platform that's trying to [17:44] meet lots of different business needs crosses  these boundaries. And one of the most valuable [17:48] lessons we learned was you really need to align  leadership to this different way of thinking. [17:53] So we've seen much greater success when the  platform owner is accountable for shared [17:56] contracts, runtime experience, uh architectural  standards, but then these capability owners sit [18:02] much closer to each business domain and they  are accountable for the purpose, the data, [18:06] the evaluation, the human review, right? The  responsible domain use in those areas. Yeah, [18:12] they they have a very narrow focus. They don't  really care about the other business domains, [18:15] right? But it's important that you have people  that are very much considering the entire thing. [18:20] The composition also needs policy. Which handoffs  are approved, which inited uses are allowed. Um [18:27] especially in a regulated industry like the  pharmaceutical industry and at Gilead where we [18:31] care a lot about um responsible AI use, it's  very important that there abundant safeguards, [18:37] controls, and processes to make sure  that we de deliver value and not harm. [18:42] So you need the funding and support  model to reflect that shared value. [18:46] You can't have funding for one capability, funding  for another capability because a lot most of the [18:50] work is actually at that platform level. And so  if anyone is embarking on on this at a much more [18:57] established enterprise with more traditional  ways of thinking, I I really recommend uh [19:01] addressing that as early as possible with  leadership. So moving on kind of the last [19:07] summarizing of everything. So the goal is not  more agents, tools, features or data. Right? More [19:11] components without trust and reliability, they  simply create more complexity and more problems. [19:17] So specialization, determinism and  durability, they create trust and reliability. [19:23] The composition and being able to combine these  things together that creates additional value. [19:27] And the architecture and the organizational  um uh system in you know at your company [19:33] that determines whether or not you can achieve  both. So let AI really re reason broadly across [19:38] the evidence and intent and figuring out what  available capabilities to use, but put complex [19:43] and exacting action inside durable workflows  that are fit for purpose and connect them through [19:49] stable contracts so that they you don't have to  constantly go back and fix and update and improve. [19:56] This is basically how multimodal platforms can  become more useful than they become complex. [20:02] And I think the timer just hit zero. But I am  happy to answer question. It's maybe offline.