Polyrific CEO Urges Kill Switches as Single AI Prompts Spawn 60 Agents

The Ungoverned Agent: Deploying Autonomous AI Without Losing Control With Polyrific — Ai4

Ai4

Enterprises face severe financial and operational exposure unless they enforce centralized identity tracking and emergency kill switches for cascading autonomous agent networks.

AI-assisted summary of The Ungoverned Agent: Deploying Autonomous AI Without Losing Control With Polyrific — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 18, 2026, this presentation features Polyrific CEO Matt Cashatt. The page description outlines how enterprises are rapidly deploying autonomous AI agents without consistent guardrails, presenting a practical infrastructure framework for real-time observability, policy enforcement, and governance at scale.

## TL;DR Matt Cashatt argues that enterprises must treat autonomous AI agents as unpriced operational risks requiring the same rigorous containment standards as nuclear power or commercial aviation. He illustrates how a single customer churn prompt can cascade into 61 autonomous sub-agents, demanding dedicated service credentials, continuous visibility, and sub-30-second emergency kill switches. Highlighting Polyrific's Catalyst platform and insurance underwriting use cases, Cashatt shows how an ambient governance layer preserves institutional knowledge while automating complex evaluation workflows. He stresses that while agents can deliver high-value recommendations, deterministic human sign-off must remain non-negotiable for critical enterprise decisions.

## Atmosphere & style The presentation blends pragmatic urgency with self-deprecating humor, creating the feel of an intimate, no-nonsense technical briefing. Cashatt balances lighthearted onstage asides—such as joking about being too old for a CIA booth or noticing the teleprompter tracking his ad-libs—with stark warnings about runaway $20,000 compute loops and nuclear-style system cascades. The tone is refreshingly grounded, puncturing typical AI industry hype to focus on the unglamorous mechanics of audit ledgers, identity badging, and practitioner empathy.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) Organizations must treat agentic risk like an insurance risk—measuring, governing, and pricing the exposure rather than ignoring it—to deploy autonomous AI safely and aggressively. - [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) Conversational interfaces like Claude or ChatGPT are merely front-end veneers that can silently trigger complex, cascading multi-agent networks behind the scenes. - [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) A single prompt can rapidly spawn dozens of autonomous sub-agents—such as 61 agents in an account churn inquiry—each requiring discrete authorization, credentials, and audit tracking. - [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) High-consequence technologies like nuclear energy and aviation rely on rigorous containment structures, instrumentation, and centralized control rooms—principles that must be applied to autonomous agents. - [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Effective governance requires five core diagnostic capabilities: identifying all active agents, enforcing separate machine identities, inspecting real-time actions, bounding failure blast radiuses, and having a sub-30-second kill switch. - [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) Comprehensive governance engines like Catalyst must maintain an exhaustive, regulator-grade operational ledger tracking agent identities, model calls, latency, costs, reasoning chains, and boundary controls. - [12:01](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=721s) While 90% of business processes can be handled by deterministic code, the remaining 10% requiring complex human judgment accounts for 90% of total enterprise value. - [12:01](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=721s) Agentic systems can preserve irreplaceable institutional knowledge from retiring senior experts to onboard junior staff, while ensuring final critical decisions strictly remain with humans. - [14:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=848s) Enterprises must catalog all shadow AI, mandate individual service credentials instead of borrowing human logins, and eliminate approval-prompt fatigue that leads workers to blindly authorize critical exposures. - [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) Runaway, unmonitored agent loops can burn through $10,000 to $20,000 in compute and API costs in minutes, making a rapid emergency 'scram button' vital. - [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) Immediate aggregation of agent logs into a single centralized repository is essential, as an imperfect or 'ugly' log pipeline right now is far better than absent visibility. - [17:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1050s) AI practitioners must recognize their insulated bubble and take responsibility for deploying technology that empowers and includes workers rather than alienating them.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) The Reality of Unpriced AI Risk in the Enterprise

Matt Cashatt, founder and CEO of Polyrific, opens the session by addressing the growing challenge of ungoverned AI agents. While Polyrific builds agentic governance platforms that apply across any domain, he draws heavily on examples from the insurance industry—a sector they partner with closely. Joking about his age, Cashatt quips that after stopping by the CIA booth at the event and finding it cool, he tried to get in, only to be told, "Nah, dude. You're old. It's not happening." Shifting to his background, he explains that Polyrific has spent 12 to 13 years building custom software and enterprise agentic solutions, grappling with how to deploy AI safely without walking into severe operational danger.

To gauge the room, Cashatt asks for a show of hands on who worries about losing track of AI agents running in production; all ten attendees raise their hands. Asking the audience to imagine a typical Tuesday one year from now, he predicts that everyone expects to have more AI woven into their professional and personal routines, not less. Turning to an audience member named John from the insurance sector, Cashatt references the concept of "unpriced risk"—phenomena that cannot be accurately valued. In insurance, companies do not automatically ignore or write unpriced risks; first, they determine the magnitude of the exposure. Cashatt argues that the same rule applies to autonomous agents: organizations must measure, govern, and price the risk of agents operating across their environments so they can confidently deploy powerful technology to win new business.

- Polyrific has operated for 12–13 years designing custom software and agent architectures. - 100% of the session room attendees indicated concern over tracking production AI agents. - Core thesis: Treating agentic risk like an insurance risk—measuring and pricing it rather than ignoring it—allows enterprises to deploy AI aggressively and safely.

---

[02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) The 61-Agent Cascade: Anatomy of a Single Prompt

Recounting a recent meeting in Houston with a long-time enterprise client, Cashatt describes discussing the rollout of large-scale agentic systems. When Cashatt emphasized the need to implement Polyrific's governed AI platform to avert potential disaster, an IT director crossed his arms dismissively and declared, "We're never going to have a situation where we're spawning 50 agents." Cashatt contends that this skepticism stems from a widespread misconception: people conflate multi-agent systems with conversational front-ends like ChatGPT or Claude. Those interfaces merely act as the user-facing veneer. Behind the scenes, a single user request can trigger a cascading tree of dozens or hundreds of autonomous agents making independent decisions.

Cashatt steps through an illustrative scenario triggered by an executive prompt: *"Hey, this account's churning. I want to know why."* The primary orchestrator agent initiates the investigation by calling sub-agents to inspect account history, support tickets, recorded calls, and past meetings. As initial findings return, those agents autonomously spawn more specialized sub-agents to identify root causes and check historical precedents. When reaching the response drafting phase, the assigned drafting agent spawns child agents to outline the narrative, enforce data leakage prevention and intellectual property safeguards, and audit the output. Before the system finishes processing that single initial prompt, the tree explodes into 61 distinct agents running across the infrastructure—each requiring its own access authority, data credentials, and audit logging.

```mermaid flowchart TD accTitle: Single Prompt Multi-Agent Spawning Tree accDescription: Flowchart illustrating how an angry CEO prompt cascades from an orchestrator into multiple tiers of sub-agents, totaling 61 agents.

subgraph Input ["User Request"] P["Prompt: Angry CEO asks why account is churning"] end

subgraph OrchestratorLayer ["Tier 1: Coordination"] O["Orchestrator Agent (Interface / Entry)"] end

subgraph Tier2 ["Tier 2: Investigation & Discovery"] AH["Account History Agent"] TH["Ticket History Agent"] CM["Calls & Meetings Analyzer Agent"] end

subgraph Tier3 ["Tier 3: Specialized Task Agents"] RC["Root Cause Agent"] PR["Precedent Analysis Agent"] RD["Response Draft Agent"] OT["Other Task Agents..."] end

subgraph Tier4 ["Tier 4: Sub-Agent Spawning (61 Total)"] OUT["Outline Agent"] NAR["Narrative Agent"] DLP["IP / Leakage Guard Agent"] AUD["Audit Agent"] SUB["Additional Sub-Agents..."] end

P --> O O --> AH O --> TH O --> CM AH & TH & CM --> RC AH & TH & CM --> PR AH & TH & CM --> RD AH & TH & CM --> OT RD --> OUT RD --> NAR RD --> DLP RD --> AUD OT --> SUB ```

- Misconception: Viewing Claude or ChatGPT as individual agents rather than entry interfaces coordinating vast agent networks. - Cascade metric: A single customer churn inquiry spawned 61 active autonomous agents across multiple hierarchical tiers. - Security requirement: Every dynamically generated agent needs bounded authorization, individual access credentials, and discrete audit tracking.

---

[06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Lessons from the Nuclear Industry: Control Rooms and Guardrails

Cashatt cues up an archival film clip depicting nuclear fission: when a sufficient quantity of U-235 reaches critical mass, ejected neutrons bombard neighboring uranium atoms, unleashing an exponential chain reaction where over a "million billion billion atoms" split within two seconds. Although emphasizing that he does not mean to trivialize real-world catastrophes, Cashatt points out that nuclear meltdowns were not failures of physics; the laws of nature remained constant. Instead, disasters resulted from human complacency, inadequate preparation, and a breakdown in system control.

Because walking away from nuclear energy was untenable amid modern power needs, society did not abandon the technology. Instead, the nuclear power sector engineered rigorous containment structures, fuel ledgers, instrumentation gauges, and centralized control rooms. Cockpits in commercial aviation operate under identical logic: systems carrying severe consequences demand dense instrumentation. Enterprises deploying autonomous agents must apply those exact containment principles. Cashatt introduces five diagnostic audit questions for technology leaders to grade their own readiness:

1. Can you name every agent that ran across your organization yesterday? 2. Does each agent possess its own distinct identity, or is it borrowing human user credentials? 3. Can you inspect precisely what any given agent is doing right now in real time? 4. If an agent experiences a failure or hallucinates, does the blast radius remain small and contained? 5. Can your team reliably trigger a full kill switch to shut down any rogue agent within 10, 15, or 30 seconds?

---

[09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) Inside Production Governance: The Underwriter View vs. The Catalyst Engine

Cashatt demonstrates how safe agent orchestration works inside a commercial insurance carrier using Polyrific's application, Advisory for Insurance. When an insurance submission package arrives containing policy applications, building inspection records, loss runs, and supplementary attachments, agents break down the documentation. They assess the submission against underwriting appetite, flag incomplete records back to the broker, score risks, and offer clear recommendations. To the human underwriter, the interface feels seamless, intuitive, and supportive rather than intimidating: "The underwriters who use this don't think of this as an AI project. It doesn't look scary. It feels helpful."

Behind this front-end interface sits Catalyst, Polyrific's agent governance engine. While the underwriter reviews business recommendations, Catalyst maintains an exhaustive, regulator-grade operational ledger tracking every agent interaction:

| Governance Metric Logged by Catalyst | Operational Scope & Purpose | |:--- |:--- | | **Agent Identity & Hierarchy** | Tracks individual agent IDs, parent callers, and spawned sub-agents | | **Model Invocations** | Details specific LLM foundational models called and third-party model providers used | | **Execution Metrics** | Captures execution runtimes, processing latency, and discrete API/token costs | | **Reasoning & Verification Trails** | Logs full chain-of-thought rationale, factual conclusions, and cross-agent peer reviews | | **Lifecycle & Boundary Control** | Enforces policy boundaries, containerized execution permissions, and termination triggers |

Cashatt warns that partial visibility does not equal governance: "If your systems at your place of work right now aren't giving you this level of specificity as to what's going on with your agents, then it's ungoverned. You can have some of it, but it doesn't mean you're governed. You need all of it."

---

[12:01](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=721s) The 90/10 Rule and Preserving Institutional Judgment

Cashatt reflects on a session from the previous day arguing that 90% of operational workflows in insurance do not require agentic AI. He agrees with the premise, reminding the audience that three decades of deterministic code already handle structured workflows efficiently without generative tools. However, dismissing AI based on that metric draws the wrong conclusion: the remaining 10% involves complex contextual judgment, and that 10% accounts for 90% of total enterprise business value.

This judgment layer is under acute pressure due to an "aging out" crisis. Experienced underwriters and senior specialists with 30 to 35 years of domain expertise want to retire but are stuck because institutional knowledge has not transferred to junior staff. By modeling expert evaluation criteria within agentic systems, companies can preserve institutional knowledge, accelerate onboarding for junior underwriters, and safeguard critical decision-making.

Crucially, Cashatt stresses that the final decision pillar must never migrate to autonomous systems. Bound or quote clearance, formal rejections, and underwriting exception grants must remain strictly with human professionals; agents should provide structured recommendations, not unmonitored executive action.

---

[14:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=848s) Practical Directives: Inventory, Identity Badging, and the Kill Switch

To conclude this part of the presentation, Cashatt outlines foundational actions that organizations can implement immediately without purchasing new software:

- **Conduct a comprehensive agent census:** Catalog every agent and agentic workflow running across the organization. Account specifically for unsanctioned tools; blocking enterprise access to Claude or ChatGPT simply drives workers into unmonitored "shadow AI" using personal accounts, wiping out organizational observability. - **Enforce individual service badging:** Mandate distinct identity credentials and dedicated service accounts for every single agent. Prohibit shared or borrowed human credentials entirely. - **Eliminate approval-prompt fatigue:** When agents borrow human accounts and repeatedly pepper workers with rapid permission pop-ups to advance a task, humans instinctively hit "yes, yes, yes" to clear the screen, eventually authorizing critical data exposures buried in blocks of text. - **Establish rapid kill switches:** Ensure technical teams maintain a dependable, sub-30-second kill procedure capable of halting any runaway agent execution instantly.

[15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) Emergency Kill Switches and Centralized Observability

Managing agentic workflows requires protecting human operators from alert fatigue while keeping catastrophic risks in check. Drawing an analogy to the nuclear power industry, Matt Cashatt highlights the necessity of a "scram button"—an emergency mechanism designed to shut down operations instantly. Within the Polyrific Catalyst platform, this exists as a literal big red button capable of halting all autonomous agent activity across an entire enterprise within seconds. Cashatt notes that catastrophic failures almost always manifest as runaway financial costs, sharing that unmonitored agentic chain reactions can easily burn through $10,000, $15,000, or $20,000 in a matter of minutes if left unchecked.

Beyond emergency controls, consolidating audit trails is the foundational first step for any organization deploying agents. Cashatt emphasizes the practical rule of thumb that "ugly beats absent"—meaning teams do not need a flawless, beautiful data pipeline in place this quarter, but they must at least aggregate their disparate agent logs into a single accessible location. As AI regulatory oversight tightens across industries, having a unified trail of agent actions is rapidly transitioning from a best practice to a legal necessity. Cashatt then pauses with a brief laugh to acknowledge an onstage realization: the teleprompter has been dynamically scrolling along with his voice the whole time, making him realize he had been ad-libbing without needing to.

- **The Scram Button:** A centralized emergency kill switch that halts all autonomous agents across an enterprise within seconds during critical failures. - **Financial Risk:** Runaway agent execution loops can incur between $10,000 and $20,000 in compute and API costs in minutes. - **Log Aggregation:** Establishing an "ugly" centralized log repository immediately is far superior to waiting on a polished logging pipeline while agents operate unmonitored.

```mermaid flowchart TD accTitle: Catalyst Agent Governance Control Architecture subgraph Execution["Agentic Execution"] A["Autonomous Agents"] --> B{"Catalyst Governance Layer"} end subgraph Controls["Governance & Safeguards"] B --> C["Real-Time Spend Tracking"] B --> D["Policy Enforcement"] B --> E["Knowledge Preservation"] B --> F["Scram Button / Emergency Stop"] end subgraph Outcomes["Resolution & Auditing"] C --> G["Centralized Audit Logs"] D --> H["Human-in-the-Loop Escalation"] F --> I["Global Process Freeze"] end ```

[16:51](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1011s) Deploying Catalyst as a Scalable Governance Layer

Cashatt argues that organizations must approach agent governance as a continuous foundational platform layer rather than as a disjointed, one-off project. Reassuring the audience, he stresses that teams should not feel behind or embarrassed if they have not implemented comprehensive guardrails yet. Autonomous agent technology is moving exceptionally fast, and every engineering team started at zero in the very recent past.

Polyrific built Catalyst specifically to rein in these ungoverned agents across enterprise environments. By treating governance as an ambient infrastructure layer, the platform enables organizations to build and scale agents safely on top of standardized controls. Core capabilities include real-time agent observability, proactive spend tracking, knowledge preservation—which Cashatt highlights as a critical operational asset—automated policy enforcement, alert routing, and seamless escalation to human-in-the-loop oversight whenever an agent encounters an edge case or high-risk decision.

- **Governance as a Layer:** Implemented as shared enterprise infrastructure rather than built bespoke for individual agent projects. - **Knowledge Preservation:** Safeguards institutional context, system interactions, and learned heuristics across model migrations and workflow iterations. - **Human Escalation:** Automatically routes low-confidence or high-impact actions to human operators before execution.

[17:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1050s) Breaking Out of the AI Practitioner Bubble

Wrapping up his session, Cashatt previews his upcoming presentation the following day, joking that he hopes at least 15 people attend. He addresses an underlying challenge facing the broader technology ecosystem: AI practitioners frequently live inside an insulated bubble where enthusiasm for the technology obscures the friction and skepticism felt by the general public. While he personally loves AI and believes in its transformative potential for society, he acknowledges that the current reality is messy and fraught with cultural and societal tensions.

Many workers and citizens currently feel alienated or disadvantaged by AI, often because they are given no voice or agency in how autonomous tools are deployed around them. Cashatt calls upon technical leaders and engineers at the "tip of the spear" to take deliberate responsibility for making AI broadly beneficial rather than concentrating its value among a select few. He concludes with lighthearted banter thanking the audience for their time and quipping about whether any of them managed to make it into the CIA, closing out the session on agent governance.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) Nah, dude. You're old. It's not happening. — Matt Cashatt

> [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) We're never going to have a situation where we're spawning 50 agents. — IT Director

> [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) Hey, this account's churning. I want to know why. — Matt Cashatt

> [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) The underwriters who use this don't think of this as an AI project. It doesn't look scary. It feels helpful. — Matt Cashatt

> [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) If your systems at your place of work right now aren't giving you this level of specificity as to what's going on with your agents, then it's ungoverned. You can have some of it, but it doesn't mean you're governed. You need all of it. — Matt Cashatt

> [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Can you name every agent that ran across your organization yesterday? — Matt Cashatt

> [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Does each agent possess its own distinct identity, or is it borrowing human user credentials? — Matt Cashatt

> [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Can you inspect precisely what any given agent is doing right now in real time? — Matt Cashatt

> [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) If an agent experiences a failure or hallucinates, does the blast radius remain small and contained? — Matt Cashatt

> [06:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=368s) Can your team reliably trigger a full kill switch to shut down any rogue agent within 10, 15, or 30 seconds? — Matt Cashatt

> [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) ugly beats absent — Matt Cashatt

## Mentioned Matt Cashatt, Polyrific, Ai4, CIA, ChatGPT, Claude, Advisory for Insurance, Catalyst

## Video description Matt Cashatt - CEO, Polyrific

Enterprises are racing to deploy AI agents that reason across systems, call APIs, and act on sensitive data, often faster than anyone can govern them. The result is a quiet crisis: thousands of autonomous agents operating without consistent guardrails, audit trails, or accountability. This session lays out the infrastructure required to deploy agents safely at scale, from policy enforcement and real-time observability to secure execution on your own data. Drawing on more than a decade building AI inside organizations that cannot afford to get it wrong, Matt Cashatt shows what an agent governance layer actually looks like in production, and why governance is fast becoming the deciding factor in whether enterprise AI delivers value or liability. Attendees will leave with a practical framework for scaling agents they can trust.

[Source](https://www.youtube.com/watch?v=a9fgDgH3nuw)

中文简报

单条AI提示词可衍生60个智能体，Polyrific CEO呼吁设立紧急中止开关企业若不对级联自主智能体网络实施集中身份追踪与紧急中止开关，将面临严重的财务及运营风险敞口。 Matt Cashatt认为，企业必须将自主AI智能体视为未被定价的运营风险，并对其采取与核电或商业航空同等严格的遏制标准。他举例说明单条关于客户流失的提示词如何级联生成61个自主子智能体，进而要求配备专用服务凭据、持续的可视性以及30秒以内的紧急中止开关。结合Polyrific的Catalyst平台及保险核保用例，Cashatt展示了泛在治理层如何在自动化复杂评估工作流的同时保留机构知识。他强调，尽管智能体能够提供高价值建议，但在关键的企业决策中，确定性的人工核准必须是不容妥协的底线。企业应像对待保险风险一样衡量和定价自主 AI Agent 的风险敞口，唯有明确评估潜在威胁并建立治理机制，才能在保障业务安全的前提下大胆推进 AI 落地。 (0:04) 对话界面仅是表层入口，单个业务 Prompt 即可在后台触发庞大的多 Agent 级联网络；例如一次客户流失调查便能自主派生出 61 个需要离散授权与审计跟踪的独立子 Agent。 (2:30) 借鉴核电与航空工业的控制室机制，企业需建立完备的围阻体系，包括为每个 Agent 赋予独立机器凭证、实时监控行为边界，并配备可在 30 秒内切断失控 Agent 的紧急停机按钮（Scram Button）。 (6:08) 虽然 90% 的常规流程可由确定性代码搞定，但需要复杂人类判断的 10% 流程构成了企业 90% 的核心价值；企业应借助 Agent 沉淀即将退休专家的隐性经验，同时将最终决策权严格保留在人类手中。 (12:01) 治理应作为企业的共享基础设施层而非一次性项目，通过建立监管级操作日志来防范失控级联，避免未经管控的 Agent 循环在几分钟内产生高达上万美元的意外算力与 API 账单。 (15:47)

Reading summary

Matt Cashatt, founder and CEO of Polyrific, argues that enterprises must treat autonomous AI agents as unpriced operational risks requiring strict, containment-level governance, rather than viewing conversational interfaces as simple, isolated tools. The Multi-Agent Cascades Behind Single Prompts (2:30) Cashatt explains that enterprise leaders frequently underestimate autonomous systems by conflating multi-agent architectures with conversational front-ends like ChatGPT or Claude. In reality, a single conversational interface serves merely as an orchestrator that can trigger vast, unmonitored agent trees. In one enterprise example, an executive prompt inquiring why an account was churning prompted an orchestrator to deploy sub-agents across ticket records, call logs, and account histories. These sub-agents then spawned further specialized child agents for root-cause analysis, IP leakage protection, narrative drafting, and output auditing, culminating in 61 distinct autonomous agents executing tasks from that single inquiry. Because each dynamic agent requires its own data credentials, execution boundaries, and audit logging, these cascades present severe operational vulnerabilities. When organizations fail to provide dedicated service identities, agents borrow human credentials and generate relentless permission pop-ups. Cashatt warns that this triggers approval fatigue, causing human workers to reflexively authorize requests without realizing what enterprise data or tools the agent is accessing. Containment Safeguards and the Scram Button (6:08) Drawing an analogy to the nuclear energy sector, Cashatt argues that high-impact technologies demand rigorous instrumentation rather than retreat. Just as uncontrolled nuclear fission leads to exponential chain reactions, unmonitored agent execution loops can spiral quickly, burning between $10,000 and $20,000 in compute and API costs within minutes. Historical nuclear accidents were not failures of physics, Cashatt observes, but breakdowns in human control and preparation that required engineered containment rooms and strict operational ledgers. To gauge operational readiness, Cashatt poses five diagnostic questions for technical leadership: whether they can identify every agent active yesterday, verify each holds an independent identity rather than human credentials, inspect agent activities in real time, isolate failures to ensure small blast radiuses, and reliably terminate rogue executions within 30 seconds. To manage extreme failure modes, he argues enterprises must maintain an emergency 'scram button'—a centralized kill switch capable of halting all autonomous agent activity across an organization in seconds. Full Observability and Log Consolidation (9:52) Cashatt stresses that partial visibility does not constitute governance. In production deployments using Polyrific’s Catalyst engine, safe orchestration requires an ambient, regulator-grade operational ledger operating continuously beneath the human interface. This architecture tracks individual agent identifiers and parent-child hierarchies, foundational model calls, execution runtimes, token costs, chain-of-thought reasoning trails, and containerized boundary policies. Without complete capture across these vectors, systems remain fundamentally ungoverned. For engineering teams that have not yet implemented comprehensive governance platforms, Cashatt advises that 'ugly beats absent.' Rather than delaying until a fully polished logging pipeline is designed, teams should immediately aggregate disparate agent logs into a single centralized repository. As regulatory scrutiny tightens around enterprise AI, maintaining a consolidated, accessible audit trail transitions from an internal best practice to a legal necessity. The 90/10 Rule and Preserving Human Authority (12:01) Addressing operational scope, Cashatt notes that deterministic software already handles roughly 90% of structured enterprise workflows without generative models. However, the remaining 10% of workflows involves complex contextual judgment, and that 10% accounts for 90% of enterprise business value. In sectors like commercial insurance, this judgment layer is acutely threatened by an aging workforce, as senior experts with decades of institutional knowledge approach retirement without transferring their heuristics to junior staff. Modeling these evaluation criteria into governed agentic workflows helps preserve institutional memory and onboard junior personnel faster. Crucially, Cashatt maintains that autonomous agents must never be granted executive decision-making power. In high-stakes environments like underwriting, bound clearance, formal policy rejections, and coverage exception grants must remain strictly under human control. Agents should be limited to breaking down documentation, flagging missing information, and presenting structured recommendations for human review. Cashatt cautions that AI practitioners often live within an enthusiastic bubble that obscures public friction, skepticism, and alienation regarding automation. He challenges engineers and technical leaders at the forefront of AI deployment to take active responsibility for building transparent guardrails and human escalation paths, ensuring autonomous tools empower workers rather than concentrating authority without accountability.

精炼阅读摘要

Polyrific 创始人兼首席执行官 Matt Cashatt 指出，企业部署自主 AI 智能体时面临未定价的运作风险，必须参照核能安全与保险定价逻辑，构建具备独立身份、全链路审计与紧急熔断的治理基础设施，以防失控级联与高昂损失。提示词级联与智能体爆炸风险 (2:30) Cashatt 认为业界普遍存在误区，容易将 ChatGPT 或 Claude 等前端对话界面等同于单个智能体，而忽视了底层由单个请求触发的多层级自主网络。他举例说明，某企业高管仅输入一条关于“客户流失原因”的简单查询，负责协调的主智能体便调用了针对账户历史、工单和会议纪要的子智能体；这些子任务随后进一步分化出归因、判例分析、内容起草、数据防泄漏及合规审计等细分智能体，最终在系统后台动态生成了 61 个独立运行的智能体。每一个动态生成的智能体都需要明确的访问权限、凭证和审计日志，若缺乏集中监控，企业在不知不觉中就会面临严重的失控隐患。借鉴核电与航空的仪表化管控 (6:08) 通过类比核裂变链式反应与商用航空驾驶舱，Cashatt 强调严重事故通常不是物理定律或技术原理失效，而是源自人类麻痹大意、准备不足和控制系统崩溃。由于现代社会无法放弃核心技术，核电行业建立了高密度的仪表中控室与遏制结构，企业部署自主智能体也必须遵循同样的严密遏制原则。他向技术管理者提出了五项诊断性审计自查标准：能否明确列举昨日组织内运行的所有智能体；每个智能体是否拥有独立身份而非盗用人类凭证；能否实时查看任意智能体的具体动作；在出现幻觉或故障时爆炸半径是否可控；以及团队是否具备在 10 至 30 秒内可靠关停失控智能体的紧急熔断能力。承保实践与 Catalyst 治理引擎 (9:52) 在保险行业的实际落地中，Polyrific 的 Advisory for Insurance 平台允许智能体解析保单申请、巡检记录及损失清单，并在后台评估承保意向与标记缺陷，为核保人员提供无缝的操作体验。然而，支撑前台的是底层的 Catalyst 治理引擎。该引擎负责维护监管级的操作日志，详细记录智能体唯一 ID 与调用层级、调用的具体大语言模型基础底座与第三方提供商、运行延迟与 Token API 财务成本、完整的思维链推导及同行交叉验证路径，并强制执行容器化边界与终止触发器。Cashatt 警示，局部可见性不等于合规治理，唯有具备全指标的透明度才能支撑稳健部署。 90/10 法则与留存人类专业判断 (12:01) 针对“保险业 90% 的流程无需智能体”的观点，Cashatt 认为虽然三十年来的确定性代码已足以高效处理大量结构化工作流，但剩下的 10% 涉及复杂的上下文判断，却占据了整个企业 90% 的业务价值。随着具备三三十余年经验的资深专家面临退休潮，行业遭遇了严重的知识断层危机。通过在智能体系统中对专家的评估逻辑与启发式思维进行建模，企业能够留存核心制度经验并加速初级员工培训。不过，Cashatt 特别强调最终决策权不可移交给自动化系统，出单核准、正式拒保和例外审批等关键底线必须始终由人类专业人员掌控，智能体仅能输出结构化建议。应急熔断机制与即时治理措施 (14:08) 针对企业即刻落地的风控需求，Cashatt 主张在无需采购新软件的情况下推行基础动作：对组织内所有智能体进行全面普查，避免因盲目封禁官方工具导致员工转向无法监控的“影子 AI”；强制实行服务身份凭证制度，杜绝借用人类账号并由此引发的“弹窗批准疲劳”；同时建立统一日志库，强调“粗糙的汇总也远胜于完全缺失”。面对不可控的失控循环，Catalyst 配备了类似于核电站的紧急停堆按钮（Scram Button），可在数秒内全局关停所有自主进程，防止智能体在几分钟内因调用死循环空耗 10,000 至 20,000 美元的算力与 API 费用。 Cashatt 提醒技术带头人走出从业者内部的技术狂热孤岛，正视公众对 AI 的摩擦与失落情绪。尽管将治理作为常态化基础设施层可防范操作与成本失控，但若不赋予普通工作者对自主工具部署的话语权，技术成果将难以公平普惠，这一文化与组织层面的挑战仍待行业深入化解。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 18 日发布在 YouTube 平台上。视频记录了 Polyrific 首席执行官 Matt Cashatt 的主题演讲，主要探讨企业在竞相部署自主 AI 智能体时所面临的失控危机，并阐述如何在生产环境中构建涵盖策略执行、实时可观测性与安全防护的智能体治理层基础设施。 ## TL;DR Matt Cashatt 在演讲中指出，自主 AI 智能体在企业内部容易形成“未定价风险”，单个简单业务 Prompt 甚至会在转瞬之间裂变出多达 61 个自主子节点。他借用核裂变链式反应的隐喻，提出企业必须建立具备实时可观测性、独立身份认证与物理级 SCRAM 熔断机制的“控制室”，防范数分钟内烧掉数万美元的失控调用。演讲还结合 Polyrific 旗下 Advisory 与 Catalyst 产品的实战经验，强调企业必须守住 90/10 价值法则中人类最终签批的底线，并将治理作为底层基础设施而非临时项目推进。 ## Atmosphere & style 演讲整体呈现出一种兼具行业沉淀与风趣自嘲的沉浸感。Cashatt 善用现场互动、CIA 展位调侃与提词器偶发的搞笑插曲来活跃气氛，但在切入核反应堆失控爆炸的影像与惨痛账单教训时，语调又陡然变得严肃警醒。整场分享节奏明快、张弛有度，既有直击工程痛点的架构拆解，又带有呼吁从业者跳出信息茧房、兼顾社会责任的人文温度。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) 未定价风险（Unpriced Risk）是企业引入 AI 的首要隐患：在无法量化评估潜在破坏量级时盲目部署自主智能体，将给企业带来无法承受的潜在负债。 - [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) 单条业务指令可引发智能体指数级裂变：一个看似简单的排查请求即可在后台自主派生出多达 61 个互相协作的子智能体，对数据访问权限与全链路审计提出严峻挑战。 - [06:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=360s) 核反应堆链式反应隐喻：未受治理的自主智能体网络如同缺乏控制机制的核裂变，企业必须建立如同核电站控制室般的安全监控与护栏机制。 - [06:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=360s) “控制室五问”治理自评清单：企业应从智能体全量盘点、独立服务身份鉴别、实时可观测性、爆炸半径限制及紧急熔断机制五个维度自检安全水位。 - [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) 精细化可观测性是合规底线：系统后台必须具备对底层模型调用、耗时、Token 成本、推理逻辑（Rationale）及多层级智能体校验链路的透明度量能力。 - [12:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=720s) 90/10 价值法则：传统确定性软件已能解决 90% 的标准化流程，但企业 90% 的核心商业价值恰恰源自那 10% 需要复杂主观经验研判的关键环节。 - [12:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=720s) 知识沉淀与 Human-in-the-loop 底线：利用智能体镜像沉淀即将退休专家的研判逻辑以赋能新人，但在保单签批等最终商业决断上严禁 AI 越界，必须由人类员工把关。 - [14:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=848s) 强制一机一账号并警惕权限轰炸：严格实行独立服务账号（Service Account）隔离，避免员工因频繁弹窗审批产生疲劳点击从而不慎放行高危权限。 - [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) 核级 SCRAM 紧急制动机制：借鉴核电站紧急停堆逻辑设置全局熔断按钮，可在数秒内叫停全组织智能体运作，防止失控的循环调用在数分钟内造成数万美元的经济损失。 - [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) “粗糙集中胜于彻底缺失”的合规原则：面对日益收紧的行业监管，即便缺乏优雅的流水线，也必须立即将分散在各处的智能体运行日志集中归档存储以备审计。 - [16:41](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1001s) 智能体治理是常态化基础设施：治理不应被视为一次性项目，而是应作为底层基础设施，贯穿全生命周期的可观测性、成本监控、策略执行与知识留存。 - [17:31](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1051s) 走出技术人员的信息茧房：AI 开发者必须认识到社会上大量普通人群尚未享受到技术红利甚至面临冲击，应共同推动技术切实惠及全社会。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) 未受治理的智能体与“未定价风险” Polyrific 首席执行官 Matt Cashatt 开门见山地指出，本次演讲的核心是探讨“未受治理的智能体（Ungoverned Agent）”。尽管他会结合深度合作的保险行业展开诸多具体案例，但智能体治理平台的底层逻辑适用于所有行业。Cashatt 幽默地自嘲道，自己不是中情局（CIA）特工——虽然会场里 CIA 的展位非常酷炫，但他去应聘时被对方以“太老了”为由拒之门外；实际上，他是一位深耕企业级软件与智能体解决方案达 12、13 年之久的创业者，长期致力于帮助企业在不陷入巨大麻烦的前提下安全落地 AI。为了切入主题，Cashatt 向现场观众提问：“在座有谁在工作中担心过无法追踪生产环境中的 AI 智能体？”台下观众一致举手。Cashatt 预测，无论是工作还是日常生活，一年后每个人身边的 AI 数量只会多不会少。他借用保险行业的术语向现场一位名叫 John 的观众互动：对于无法量化标价的潜在危机，精算上称之为“未定价风险（Unpriced Risk）”。企业面对未定价风险既不能视而不见，也不能盲目承保，而是必须先搞清其风险量级。Cashatt 强调，掌握对 AI 智能体风险的定价与管理能力，正是企业在未来安全释放这项强大技术红利、拓展业务的关键前提。 - **核心行业概念**：未定价风险（Unpriced Risk）——在无法评估潜在破坏量级时盲目推进，将给企业带来无法承受的负债。 - **互动调查**：现场参会者 100% 担忧无法在生产环境中有效追踪并监管自主 AI 智能体。 --- ### [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) 一条 Prompt 引发的连锁反应：61 个智能体的诞生许多企业管理者容易陷入一种认知误区，将单一的 ChatGPT 或 Claude 聊天窗口误当成全部的 AI 智能体。Cashatt 分享了数周前在休斯敦与一家合作多年的大客户高管沟通的经历。当 Cashatt 警告必须引入 AI 治理平台以防失控时，该客户的 IT 总监抱起手臂不以为然地反驳：“我们绝不可能搞出一次性派生出 50 个智能体的情况。” 为了打破这种误解，Cashatt 现场拆解了一个看似极为简单的企业场景：一位愤怒的 CEO 输入了一条 Prompt——“这个大客户准备流失了，我要知道原因。”在系统后台，负责接收指令的主协调器（Orchestrator）立即开始派生第一梯队的智能体去分别调取账户历史、工单记录以及会议录音；当首批信息返回后，协调器又自主决定派生更多节点去分析根本原因、查找历史先例；紧接着，负责撰写回复草案的智能体（Response Draft Agent）又继续自发派生出大纲拟定、叙事构建、防 IP 泄露审查以及全流程审计等专属子智能体。层层裂变之下，仅为了响应一条简单的排查指令，后台在转瞬之间便自主派生出了多达 61 个互相协作的子智能体。 ```mermaid flowchart TD accTitle: Agent Spawning Cascade Diagram accDescription: Flow showing how a single executive prompt expands into a multi-tier hierarchy of 61 autonomous agents. subgraph UserTier ["用户输入层"] Prompt["CEO Prompt: '客户流失原因分析'"] end subgraph OrchestrationTier ["主协调与第一层级"] Orch["主协调器 Orchestrator"] AccAgent["账户历史分析智能体"] TicketAgent["工单记录智能体"] CallAgent["通话与会议解析智能体"] end subgraph DeepAnalysisTier ["深层归因与策略层"] RootCause["根本原因诊断智能体"] Precedent["历史先例匹配智能体"] DraftAgent["响应草案生成智能体"] end subgraph ExecutionSubAgents ["叶子执行节点（裂变多达 61 个）"] Outline["大纲构建子智能体"] IPProtect["数据防泄漏/IP审查子智能体"] AuditSub["合规审计子智能体"] MoreAgents["其余数十个动态派生节点..."] end Prompt --> Orch Orch --> AccAgent Orch --> TicketAgent Orch --> CallAgent AccAgent --> RootCause TicketAgent --> Precedent CallAgent --> DraftAgent DraftAgent --> Outline DraftAgent --> IPProtect DraftAgent --> AuditSub RootCause --> MoreAgents ``` - **单次请求的裂变规模**：从 1 个业务 Prompt 最终在底层自主派生出 61 个具备独立决策和数据访问需求的智能体节点。 - **权限与审计挑战**：这 61 个智能体中的每一个都需要独立的凭据去读写企业数据，并产生严密的审计日志以供合规回溯。 --- ### [06:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=360s) 核反应堆隐喻与控制室五问 Cashatt 在现场播放了一段展示铀-235（U-235）核裂变链式反应的影像资料，指出高速中子轰击原子核导致数以亿万计的原子在 2 秒内发生失控爆炸。他严肃地指出，这正是未受治理的智能体网络在企业内部可能造成的破坏隐喻。历史上核灾难的根源并非物理定律失效，而是人类在准备不足或缺乏控制机制下的操作失误。人类并没有因为潜在的核风险而放弃这项巨大的能源，而是通过建造精密的仪表盘、控制室和安全护栏来驯服它。对待 AI 智能体同样如此——企业绝不能在没有控制室的情况下盲目部署“裸奔”的自主智能体。为此，Cashatt 提出了企业必须面对的“控制室五问”自评清单： 1. **盘点能力**：如果现在被要求，你能否准确说出昨天在企业内运行的每一个智能体名称？ 2. **身份鉴别**：每个智能体是否拥有专属的独立身份标识，还是在盗用/借用人类员工的身份凭证？（借用人类凭证往往是灾难的开端） 3. **实时可观测性**：你现在能否实时看清任意一个智能体正在执行的具体操作？ 4. **爆炸半径**：一旦某个智能体发生运行偏差或异常，事故能否被限制在可控的小范围内？ 5. **熔断机制**：如果出现失控苗头，你是否能在 10 秒、15 秒或 30 秒内将其一键彻底关停（Kill Switch）？ --- ### [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) 行业实战：Advisory 承保应用与 Catalyst 治理平台 Cashatt 展示了 Polyrific 旗下的两款核心产品如何配合落地：面向业务端的是智能承保应用“Advisory for Insurance”，而后台运行的则是全托管治理平台“Catalyst”。在财产险或车险业务中，投保文件包（Submission Package）通常充斥着冗长的申请表、现场查勘报告及各类补充文件。当文件包进入系统后，Advisory 会派生多个智能体迅速拆解内容，对照保险公司的承保意愿（Underwriting Appetite）进行匹配，若信息缺失则自动触发补件流程，并给出风险评估建议。在前端核保师看来，整个界面极其友好，感觉就像一个贴心的业务助手，完全感受不到“AI 的冰冷与压迫感”。但在后端，Catalyst 治理平台对全流程进行了严密监控与度量：精确记录调用的每一个底层大模型（Model Providers）、耗时、Token 成本、推理推导逻辑（Reasoning / Rationale）、派生的所有子智能体链路，以及第二层智能体对第一层智能体的校验反馈。Cashatt 直言：“如果你在公司使用的系统无法提供这种精细度的可观测性，那你的 AI 系统本质上就是完全未受治理的。” --- ### [12:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=720s) 90/10 价值法则与不可逾越的人类底线在谈及 AI 智能体在行业中的定位时，Cashatt 引用了前一天会议上的一个观点：“保险行业 90% 的工作其实根本不需要智能体 AI。”他对此表示认同，并坦言过去 30 年间确定性软件（Deterministic Software）已经非常出色地解决了大部分标准化流程，没必要搞“AI 泛滥”。然而，很多人得出的结论却是错误的：企业真正的核心护城河与“90% 的商业价值，恰恰来源于那 10% 需要复杂主观判断（Judgment）的核心环节”。面对行业资深核保专家即将大批退休（Aging out）带来的经验断层危机，通过智能体将老专家的研判逻辑镜像化保存下来，可以极大地赋能并加速年轻员工的成长。Cashatt 明确划定了业务治理的三大边界： | 业务维度 | 技术实现方式 | 责任主体 | 业务示例 | |:--- |:--- |:--- |:--- | | **标准/常规流程 (90%)** | 传统确定性代码 / 规则引擎 | 系统自动化 | 表单解析、数据归档、格式校验 | | **经验研判与知识沉淀 (10%)** | 自主协同 AI 智能体 | 智能体推荐 + 辅助分析 | 复杂风险画像、投保意愿匹配、先例对比 | | **最终商业决断 (核心边界)** | 严禁 AI 越界，必须 Human-in-the-loop | 具备资质的人类员工 | 最终保单签批（Bind）、正式报价确认、拒保裁决 | --- ### [14:08](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=848s) 落地行动指南：智能体盘点与权限陷阱在演讲第一部分的尾声，Cashatt 给出了所有企业参会者返回工作岗位后即可立即执行的三项安全行动准则： 1. **全面清点智能体与工具**：无论官方批准与否，彻底摸排并登记企业内正在使用的所有智能体工具。简单粗暴地封禁 ChatGPT 或 Claude 只会将员工推向“影子 AI（Shadow AI）”，导致 IT 部门彻底丧失可观测性。 2. **强制服务账号与身份工牌（Badging）**：严格实行“一智能体一服务账号（Service Account）”原则，严禁借用人类员工权限。 3. **警惕“权限轰炸”与疲劳点击**：Cashatt 指出，当员工被海量的系统弹窗反复询问“是否允许执行此操作”时，往往会产生疲劳心理，甚至不看内容就连续点击“Yes”，从而在密密麻麻的文本墙中不慎放行致命的高危权限——这与核电站事故中的人为疏忽如出一辙。 4. **常备熔断切断程序（Kill Procedure）**：必须为生产环境中的每个智能体链路配备快速生效的物理级刹车机制。 ### [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) 紧急制动按钮与日志集中化：防止失控的链式反应在管理面向员工的各类 AI 系统时，企业必须时刻掌握全局，以避免人员产生“警报疲劳”与操作过载。而在安全底线方面，Matt Cashatt 提到了核工业领域著名的“紧急停堆按钮”（SCRAM button）。在核电站中，SCRAM 机制可以在极端危险时瞬间终止反应堆核心运作；借鉴这一理念，Polyrific 在其治理平台 Catalyst 中设置了一个“醒目的大红按钮”。只要企业发现任何异常或系统走向失控，管理员能够在数秒内一键切断全公司范围内的所有智能体活动。这种紧急阻断功能之所以至关重要，根本原因通常在于真金白银的损失。Matt 坦言，团队过去曾亲身经历过智能体失控的链式反应（runaway chain reactions），在短短几分钟内就凭空烧掉了 1 万、1.5 万甚至 2 万美元。面对这种指数级吞噬资源的意外，企业必须拥有一键叫停全盘运作的能力。与此同时，企业应当立即着手将分散的现有日志集中化。即便本季度还来不及搭建一套精美完备的流水线，“丑陋粗糙的集中也远胜于彻底缺席（Ugly beats absent）”。随着监管政策日益趋严，无论企业是否选择 Catalyst 这样的专业平台，都必须清楚掌握每一条日志的流向与位置，确保审计链条完备。 ```mermaid flowchart TD accTitle: 智能体异常防护与失控阻断机制 subgraph RunawayRisk["失控风险场景"] A["智能体循环调用 / 链式反应"] --> B["API 账单飙升<br/>数分钟损失上万美元"] end subgraph GovernanceIntervention["Catalyst 治理干预"] C["实时观测与日志归集"] --> D{"异常检测 / 成本阈值触发"} D -->|常规警告| E["向人机协同 (HITL) 升级告警"] D -->|灾难性失控| F["SCRAM 紧急制动按钮 (大红按钮)"] end B --> D F --> G["数秒内阻断全组织智能体执行"] ``` - **核级 SCRAM 制动**：借鉴核工业的紧急停堆逻辑，Catalyst 提供全局“大红按钮”，数秒内终止全组织所有活跃 Agent。 - **真金白银的惨痛教训**：智能体链式循环调用失控时，短短几分钟内即可产生 10,000 到 20,000 美元的云端与 API 消耗。 - **“粗糙好过缺失”的合规原则**：面对愈发严格的行业监管，哪怕缺少优雅的 Pipeline，也必须把散落各处的运行日志集中存储起来。 --- ### [16:41](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1001s) Catalyst 治理层：将智能体治理作为常态化基础设施演讲进行到此处，Matt 忽然逗趣地发现台下的提词器居然是随着自己的说话语速自动滚动的，不由得笑着自嘲道：“我才意识到提词器会跟着我说话往下滚，那我刚才白临场发挥了那么久！”言归正传，他指出企业应该将智能体治理视作底层的基础设施层（Layer），而不是当成一个一次性的零星项目（Project）来对付。对于目前尚未建立完备治理体系的团队，Matt 也送上了宽慰：智能体是一项极新的前沿技术，迭代演进极快，Polyrific 也是从零起步、历经摸索才掌握了治理脉络。为了彻底降服那些不受约束的智能体（Ungoverned Agents），Polyrific 研发了专用的智能体治理层平台 Catalyst。当企业在该平台上构建并托管智能体应用时，可以获得全生命周期的可控性： - **全方位智能体可观测性（Full agent observability）**：全面监控 Agent 的思考路径、调用行为与工具使用。 - **成本与支出追踪（Spend tracking）**：实时盯紧 Token 与 API 消耗，防范账单爆表。 - **知识资产沉淀与留存（Knowledge preservation）**：Matt 特别强调这是极核心的价值点，确保智能体在交互中沉淀的经验与业务知识不流失。 - **策略合规执行（Policy enforcement）与预警**：从底层强制推行安全防护栏与行为准则。 - **人机协同（Human-in-the-loop）升级机制**：在关键或存疑决策点及时介入人工审查。 --- ### [17:31](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1051s) 走出从业者信息茧房：让 AI 真正惠及全社会在演讲临近尾声时，Matt 特别借用最后几分钟时间，为自己次日的另一场分享做预告。他半开玩笑地表示希望明天台下至少能来 15 个人，因为那是他在前一周撰写演讲稿时深有感触、必须向业界呼吁的重要议题。 Matt 语重心长地指出，目前身处 AI 领域的从业者们大多身陷一个由狂热构筑的“信息茧房”之中。圈内人对 AI 极度热爱，他也笃信 AI 拥有造福全人类的巨大潜力；然而现实远没有那么简单——社会上有大量普通人并未从当前的 AI 浪潮中享受到红利，甚至在根本没有选择权的情况下被迫接受技术带来的冲击，这已经演变为一个不容忽视的文化与社会问题。Matt 呼吁，作为行业中最前沿的“刀锋”（tip of the spear），AI 建设者们必须团结起来深入反思，确保这项强大的技术能够切实造福所有人，而不仅仅是少数精英群体。最后，Matt 以幽默的话语向全场致谢告别：“很遗憾你们其实没能加入 CIA，或许你们已经进了，我不知道呢——非常感谢大家！” ## Notable quotes > [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) 在座有谁在工作中担心过无法追踪生产环境中的 AI 智能体？ — Matt Cashatt > [00:04](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=4s) 对于无法量化标价的潜在危机，精算上称之为‘未定价风险’。 — Matt Cashatt > [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) 我们绝不可能搞出一次性派生出 50 个智能体的情况。 — 客户 IT 总监 > [02:30](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=150s) 这个大客户准备流失了，我要知道原因。 — CEO Prompt > [06:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=360s) 历史上有核灾难的根源并非物理定律失效，而是人类在准备不足或缺乏控制机制下的操作失误。 — Matt Cashatt > [09:52](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=592s) 如果你在公司使用的系统无法提供这种精细度的可观测性，那你的 AI 系统本质上就是完全未受治理的。 — Matt Cashatt > [12:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=720s) 保险行业 90% 的工作其实根本不需要智能体 AI。 — 会议参会嘉宾 > [12:00](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=720s) 90% 的商业价值，恰恰来源于那 10% 需要复杂主观判断的核心环节。 — Matt Cashatt > [15:47](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=947s) 丑陋粗糙的集中也远胜于彻底缺席（Ugly beats absent）。 — Matt Cashatt > [16:41](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1001s) 我才意识到提词器会跟着我说话往下滚，那我刚才白临场发挥了那么久！ — Matt Cashatt > [17:31](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1051s) 目前身处 AI 领域的从业者们大多身陷一个由狂热构筑的‘信息茧房’之中。 — Matt Cashatt > [17:31](https://www.youtube.com/watch?v=a9fgDgH3nuw&t=1051s) 很遗憾你们其实没能加入 CIA，或许你们已经进了，我不知道呢——非常感谢大家！ — Matt Cashatt ## Mentioned Matt Cashatt, Polyrific, Catalyst, Advisory for Insurance, ChatGPT, Claude, CIA (中情局), Ai4 ## Video description Matt Cashatt - CEO, Polyrific Enterprises are racing to deploy AI agents that reason across systems, call APIs, and act on sensitive data, often faster than anyone can govern them. The result is a quiet crisis: thousands of autonomous agents operating without consistent guardrails, audit trails, or accountability. This session lays out the infrastructure required to deploy agents safely at scale, from policy enforcement and real-time observability to secure execution on your own data. Drawing on more than a decade building AI inside organizations that cannot afford to get it wrong, Matt Cashatt shows what an agent governance layer actually looks like in production, and why governance is fast becoming the deciding factor in whether enterprise AI delivers value or liability. Attendees will leave with a practical framework for scaling agents they can trust. [Source](https://www.youtube.com/watch?v=a9fgDgH3nuw)

Source quotes

[0:04] Matt Cashatt: 在座有谁在工作中担心过无法追踪生产环境中的 AI 智能体？ [0:04] Matt Cashatt: 对于无法量化标价的潜在危机，精算上称之为‘未定价风险’。 [2:30] 客户 IT 总监: 我们绝不可能搞出一次性派生出 50 个智能体的情况。 [2:30] CEO Prompt: 这个大客户准备流失了，我要知道原因。 [6:00] Matt Cashatt: 历史上有核灾难的根源并非物理定律失效，而是人类在准备不足或缺乏控制机制下的操作失误。 [9:52] Matt Cashatt: 如果你在公司使用的系统无法提供这种精细度的可观测性，那你的 AI 系统本质上就是完全未受治理的。 [12:00] 会议参会嘉宾: 保险行业 90% 的工作其实根本不需要智能体 AI。 [12:00] Matt Cashatt: 90% 的商业价值，恰恰来源于那 10% 需要复杂主观判断的核心环节。 [15:47] Matt Cashatt: 丑陋粗糙的集中也远胜于彻底缺席（Ugly beats absent）。 [16:41] Matt Cashatt: 我才意识到提词器会跟着我说话往下滚，那我刚才白临场发挥了那么久！ [17:31] Matt Cashatt: 目前身处 AI 领域的从业者们大多身陷一个由狂热构筑的‘信息茧房’之中。 [17:31] Matt Cashatt: 很遗憾你们其实没能加入 CIA，或许你们已经进了，我不知道呢——非常感谢大家！

Source quotes

[0:04] Matt Cashatt: Nah, dude. You're old. It's not happening. [2:30] IT Director: We're never going to have a situation where we're spawning 50 agents. [2:30] Matt Cashatt: Hey, this account's churning. I want to know why. [9:52] Matt Cashatt: The underwriters who use this don't think of this as an AI project. It doesn't look scary. It feels helpful. [9:52] Matt Cashatt: If your systems at your place of work right now aren't giving you this level of specificity as to what's going on with your agents, then it's ungoverned. You can have some of it, but it doesn't mean you're governed. You need all of it. [6:08] Matt Cashatt: Can you name every agent that ran across your organization yesterday? [6:08] Matt Cashatt: Does each agent possess its own distinct identity, or is it borrowing human user credentials? [6:08] Matt Cashatt: Can you inspect precisely what any given agent is doing right now in real time? [6:08] Matt Cashatt: If an agent experiences a failure or hallucinates, does the blast radius remain small and contained? [6:08] Matt Cashatt: Can your team reliably trigger a full kill switch to shut down any rogue agent within 10, 15, or 30 seconds? [15:47] Matt Cashatt: ugly beats absent

Original transcript

[0:04] So, we're going to talk about the ungoverned agent. And it's not really just for insurance. [0:09] The lessons are going to be transferable to everything else that that we do. Uh what we [0:13] do is is uh in terms of like an agentic governed platform. It works for all industries, but we're [0:18] going to give a lot of examples for the insurance industry, which we work closely with. Okay. So, [0:24] I'm at Gashad. I'm not with the CIA. I don't know how many of you made it by the CIA booth that was [0:28] here. We thought all thought that was super cool. I tried to get in. They're like, "Nah, [0:31] dude. You're old. It's not happening." Um, what I am though is a founder who has founded a company [0:38] that's worked with organizations for 12, 13 years building custom software, building agentic [0:44] solutions, trying to get our hands like the rest of us are around how are we going to make this [0:48] work in the workplace without getting ourselves into really, really deep trouble. So, let me do [0:53] the the quick show of hands. How many of you worry at work that you're not going to be able to track [0:57] AI agents in production? Awesome. All 10 hands in this room raised. Cool. So, um, look, it's a worry [1:06] that we all have, but picture yourself a year from now and I want you to kind of ask yourself, [1:10] is there going to be more AI in your life or less? I'm talking about your picture your work life, [1:15] your home life, just a regular Tuesday, nothing special. I'm going to make a prediction that most [1:20] of you all are predicting that you're going to have more AI in your life. So, we're going to [1:24] talk about governance today. We're going to talk about how daunting the task is. But we all have to [1:28] face it anyway. Insurance people in insurance, can I get your name, sir? John. John probably [1:35] knows what I'm talking about when I say they have a name for a phenomenon for something that they [1:40] can't put a price tag on. It's an unpriced risk. Am I getting that right? Unpriced risk. Okay. [1:45] They don't necessarily ignore those unpre price risks and they don't necessarily write them, [1:50] but would like they they like to do is find out what the magnitude of that risk is going to be [1:54] before they write it. So, here's my claim for the next 15 20 minutes or we could all just go [1:58] get a beer, whatever's more fun. There's only a few of us in here. Um, my claim for the next 15 [2:03] or 20 minutes is going to be that getting your hands around how to price that risk. Getting [2:08] your hands around how to uh um manage the risk of of agents running around your organization, [2:15] which is happening right now, um is going to allow you, if you're in the insurance world, [2:19] to price more of that risk to go ahead and and and get more business or whatever industry that you're [2:24] in. It's going to help you go forward and do more things with this very powerful technology that we [2:29] definitely don't want to ignore. This is something that happened to me. Um, actually my my company, [2:35] my team was with me in Houston a few weeks ago. We sat down with a a customer of ours that we've [2:40] worked with for several years now. They're a big account for us. They trust us. We trust them. [2:44] Great guys. Uh, we were having to talk with them about rolling out some big applications, [2:49] agentic applications across the organization. And I said, "Listen, we need to start um implementing [2:54] our governed AI platform for you um because if we don't get our hands around what these agents [2:58] are doing, it could spell disaster for you." And the IT manager, one of them, the director, [3:03] sat back with, you know, arms across his chest and was like, "We're never going to have a situation [3:06] where we're spawning 50 agents." And I said, "Oh, really? That's not the case." And this is the [3:12] problem that a lot of people are conflating and maybe a lot of the people in this room, [3:15] you conflate an AI agent with chat GPT or Claude, and that's simply not the case. That is a face. So [3:20] it's a user interface on top of many many agents and for one simple request they can spawn up to [3:25] hundreds of agents. One agent can spawn other agents. That's why they have agency, right? They [3:30] can make choices. They can go do things that you, you know, you and I might not know that they're [3:34] doing if you don't have some kind of observability platform. So let's watch what happens. So here's [3:39] this, you know, this is all mock, you know, phony stuff, but here's this this uh prompt going in. [3:45] This is some angry CEO saying, "Hey, you know, this account's churning. I want to know why." [3:49] All right. So, the first uh the first agent, the orchestrator that you always have, that's [3:54] that's the one answering your calls at chat GPT or Claude or whatever you use is saying, "Okay, [3:58] we're going to go pull the account history. We're going to go pull ticket history and then calls and [4:02] meetings and analyze those things and start seeing why is this client upset with us right now? Why do [4:07] why might they turn? Why might we lose them? And what can we do about it?" So, it gets to work, but [4:11] it quickly discovers when it gets some information back from from the first two that, hey, I actually [4:16] need to spawn a few more agents because there's more information that I need. And then it you gets [4:20] feedback there and it spawns a few more. So, this is just the first tier of agents that it spawns [4:24] to go answer the request that came from the user. And all this, if you look at it, makes sense. I [4:31] don't think I have to go through a mad nauseium. You probably can fill in the blanks in your own [4:34] mind of a situation you've had where, yeah, we got an angry customer. uh we need to understand the [4:38] root cause of what what what the problem was that drove them to be angry. Uh we need to look at the [4:42] precedent. Have we had this situation with another customer before? All of these these nodes in here, [4:47] this first node generation of agents was carefully thought out and and and designed to be relatable [4:52] for you all because I don't want to be just another tech speaker up here uh that doesn't [4:56] relate to business. All right. So now we have let's just take the response draft agent. It's [5:02] going to go in there. It's going to spawn its own agents to say, "All right, well, we need to [5:05] outline this thing. We need to build the narrative around it. We need to make sure that we we pull [5:08] out any any information that shouldn't be leaking outside of the organization, including IP that [5:12] could leak. And then we need to audit everything, right? But it's not just the response draft agent. [5:18] All of the agents are going to start spawning their own sub agents to do that job. And this will [5:23] keep happening and keep happening over and over again until There we go. Now we're going to have [5:31] more agents spawning. There we go. Now, now more agents spawning, more things and stuff happening, [5:38] and this is how how the agentic world works. So, now we're at 61 agents. And when we kind of took [5:44] our client through this, they're like, "Wait, I had no idea." I was like, "Yeah, this is how it [5:48] works." And every one of those needs to have its own um own authority, its own set of credentials [5:54] to get in and out of your data. It's its own audit log so that we know exactly what's happening. [6:00] So 61 agents, one request. You know, one more thing for you all to know. This was all based [6:08] on on actual stuff that happens. I mean, none of that was contrived or made up. I mean, if [6:11] you if you watch how a lot of your agentic system works, if you have that level of observability, [6:16] you're going to see exactly what it does. Now, I was informed when I walked in here, we don't [6:19] have sound. So, this next little video clip isn't going to have any sound. But this pattern here are [6:24] free neutrons driven out with tremendous speed. Do have sound. And provided there is sufficient U235 [6:30] present, what science calls a critical mass, those neutrons bombard other uranium atoms, [6:37] causing them to split and split still others. The result, a chain reaction over a million billion [6:46] billion atoms exploding within 2 seconds. A million billion billion. It's a big number, [6:55] right? So we have seen this before. We've seen a very similar pattern. It's the chain reaction. [7:00] And you know, unfortunately, I don't want to make light of some of the disasters that happen around [7:03] those things. But it is a useful metaphor for what we're doing now because what happened in [7:09] the in the situations where we had these runaway reactions weren't failures of physics or nature. [7:15] The laws of physics don't change, right? What did change was the preparation or the lack of [7:20] preparation or the way it was approached to manage these things. The silver lining in in all of that [7:26] is that that ungoverned approach can be governed. We can build guardrails. We can do things. So, [7:32] it's not such a daunting task. Nobody in this room is trying to build something completely [7:35] ungoverned, right? We're not building nuclear reactors without this the right guardrails. We're [7:39] not building agentic systems without the right guardrails. We just have to slow down and ask the [7:43] right questions. And you know, the interesting thing about all this is what does that physics [7:48] do when we start putting instrumentation to it? and we start putting guard rails around it. The [7:53] nuclear industry didn't stop when they had those those catastrophes, right? It got instrumented. We [8:00] built the guardrails. We made it safer because a world without harnessing such an awesome powerful [8:05] thing wasn't an option. We all knew we needed and it's a good thing because we'd be deep in [8:09] an energy crisis right now if we didn't. So, you know, you do move forward with these things even [8:14] when they scare you a little bit. We're better off for having done it. Now, sometimes the risk seems [8:21] just too big, but you move forward. You price it, John. We price it, right? If you're in the [8:26] insurance industry, you price the risk. Can you keep moving forward? Get your hands around it. So, [8:31] score yourself on a few questions here. You don't have to answer out loud. Um, if you needed to, [8:36] if you're pressed, could you go and name every agent that ran yesterday? Does each of your agents [8:42] have its own identity or is it borrowing the humans? Bad news if it's borrowing the humans. [8:48] Can you see what any one of those agents is doing right now? If one of those agents goes wrong, [8:56] does the accident stay small? And can you stop any of these in 10, 15, 30 seconds? [9:04] If we want to learn from 80 years of experience in the nuclear industry, then this is what we [9:08] should take away. We need this kind of a control room. They keep a fuel ledger. They keep they keep [9:15] things contained. They have gauges to watch what what's happening at all times. You see [9:19] this in the cockpit of the airplane too when you step on things with high consequences have more [9:23] instrumentation because we need to know what's going on. It's no different for any of you at work [9:27] right now. What might surprise some of you all and of course I don't know each and every one of you, [9:32] but some of you are going to have different levels of of experience um with AI agents and what they [9:37] can do. What you might not realize if you're not deep into this stuff like some of the big AI geeks [9:42] like myself are is just how deep they go and how many of them that you do have running all around [9:47] you at work and at home right now that you're not even aware of. It's a big number and we don't walk [9:52] away from something powerful. We just harness it. So here's what harnessing that looks like inside [9:57] an insurance carrier. This is an application of ours called Advisory for Insurance and Gentic [10:02] app. We got mock data plugged into this one. Uh that helps with underwriting and then some more of [10:07] the process in the insurance world. So submission package and insurance is just a big pack package [10:11] with application for insurance coverage for your home, for your business, your car, what have [10:15] you. Um and it includes inspection reports and all kinds of other supplemental information. So in our [10:20] world, an agent reads that when it comes through and it checks it against appetite. Does the you [10:25] know, does the agency does the carrier want to underwrite this risk? Um it it will send back [10:29] the referral if it's not complete. if the package isn't complete and say, "Hey, we're missing these [10:34] things. This is all the agents tearing this thing apart, tearing this package apart. Lots [10:37] of them getting spawned just like we showed with the animation earlier." And then coming back and [10:41] saying, you know, things are missing. It's a good risk, it's bad risk as it makes recommendations. [10:47] That's the screen that the act the underwriter in this case would actually use. But it's not [10:51] all that's going on behind the scenes. And nobody that's using it. The underwriters [10:55] use this don't think of this as an AI project. Doesn't look scary. It feels helpful. At least [10:59] that's feedback we're getting from our customers. This is what's happening behind the scenes though. [11:04] This is our agent governance platform that we call catalyst. It is measuring every every uh agent, [11:11] watching every agent, containing every agent, managing the life cycle, uh controlling what [11:15] they can do. So um this is something that you or your IT department or regulator can go see based [11:21] on that same decision. So that same submission package that we were just showing you in the [11:24] other screen in the user interface is backed by this audit trail of everything that happened. [11:29] Every agent that was used, every model that was invoked, model providers, how much, how long, [11:34] um what it cost, you know, what it concluded, the reasoning, the rationale, the sub agent spawned, [11:39] what the second agent said when the first one when it checked the first agent, all of that is [11:43] logged. If you can't do that, if if your systems at your at your place of work right now aren't [11:50] giving you this level of specificity as to what's going on with your agents, then it's ungoverned. [11:56] It's ungoverned. I mean, you can have some of it, but it doesn't mean you're governed. You [12:01] need all of it. Now, the harder question, this is one that I get from a lot of carriers when [12:05] I talk to them and other people in in different industries. Where does this work actually belong? [12:10] This came up yesterday in a talk that I attended and it was thoughtprovoking. And that was this [12:16] particular talk was about the insurance industry that 90% of the work doesn't doesn't need Agentic [12:22] AI. And you know what? They're right. You're going to hear that from time to time at this [12:27] conference and in other ones. You'll definitely hear it behind closed doors when you're meeting [12:30] with people. You know, I I'm I'm in an AI bubble. I love AI. Not everybody does. And a lot of the [12:36] problems have been solved over the past 30 years already with deterministic software. We don't need [12:41] AI for everything. But here's the little thing that I think is the wrong conclusion about that. [12:45] The 10% where you do the judgment is where 90% of the value comes from. This is what's saving [12:52] people. And it's more than just an interesting problem. We're facing this big aging out right [12:57] now of experience. So in the insurance world, you have underwriters who have been around for 30, [13:02] 35 years. A lot of them want to retire and they don't feel like they can because the experience [13:06] is not in the younger ranks yet. They want it to be there, but they need ways to preserve their [13:11] knowledge, to preserve their decision-making. And that's where the agents step in and help. So, [13:16] these are urgent problems. They're things that are coming up. You can probably imagine whatever [13:19] industry you're in where you have a similar problem where knowledge preservation is important. [13:24] You can mirror that knowledge inside AI agents and help make junior people learn faster, bring them [13:29] along faster. Uh, so it's not about getting rid of of jobs. is about lifting other people up and [13:34] making sure that you're letting the the the people who are ready to move out the door move out the [13:37] door. The third column on this this chart never moves over. Um that is that's what stays with a [13:43] human being and in my opinion that needs to be a lot. So in the insurance world whether or not [13:47] to to bind or clear for quote a policy um or a submission package rather uh the exceptions that [13:53] they might make the decisions they might make about you know whether to deny that package or [13:57] not. These are human decisions and we don't we don't uh uh take those on. The agents don't take [14:02] those on. They just make recommendations. So that was a one carrier. U here's the part that works [14:08] everywhere in any industry and none of it requires buying anything. You need to count your agents [14:14] when you get back to work. There's just a few things that you can do. Count the agents. Count [14:18] the agentic tool. Every aentic tool, sanctioned or not, whether you know about it or not, count [14:22] them up. Document it somewhere. Um, and then the unsanctioned ones are the ones you need to know [14:28] the most about because if if you've just denied access to to to be cheap here, chat GPT or claw at [14:34] work, what you've done is you've you've given the green light really to people to just go do it in [14:39] the shadow AI world. They're going to use personal tools. Now you have zero observability. You need [14:44] to badge everything. Everything needs a service account. So if you have agents, if you're in [14:48] IT here or just ask your IT department what this means. Every agent should have its own identity, [14:53] its own service account. uh one per agent, no borrowed human credentials. Let me talk real [14:57] quick about why that's important. If if you've ever seen something like this, it's very similar [15:02] to what happened in some of these run chain chain reactions at the uh the nuclear meltdowns [15:06] that we've experienced. The human being is sitting there trying to get a job done, right? And they're [15:10] asking uh the agents to build this, to build that, to get this PowerPoint done, whatever it is, [15:16] and they start getting bombarded with permission requests. Is it okay if I do this? Is it okay if [15:21] I do that? Is it okay? And before you know it, the human's just sitting there clicking yes, yes, yes, [15:24] yes, yes, yes, yes, yes. Just leave me alone, and they're not paying attention to that one critical [15:29] permission that was buried in the middle of a wall of text. This happens over and over again. So, [15:35] it's really, really important. It's really, really important that we're aware of what tools are being [15:41] used and that that you have the kill procedure. I need to not forget about that one. and and that [15:47] that you have a handle of of everything that's facing your people so they don't get fatigued. Um [15:54] the kill button in the the nuclear industry they call that a a scram button. You can look that up [15:59] on your own. It's a uh it's it's an interesting concept. But we have a big red button in Catalyst [16:04] that lets you stop all agentic activity across the organization in seconds if you need to if [16:09] something's going horribly wrong. Usually that resolves to something involving money. [16:14] We've had before runaway chain reactions that cost us 10520,000 in a matter of minutes. So [16:21] you you need to have a button where you can kind of stop everything and centralizing the logs that [16:25] you already have. Uly beats absent. Um you don't have to have a beautiful pipeline this quarter, [16:30] but make sure at least if you're not going to use a platform like Catalyst that your logs are coming [16:33] together that you know where those things are because everything's starting to become more and [16:37] more um uh regulated, which is good. I just not realized that the little teleprompter scrolls when [16:44] I talk. So, I've been ad libbying for no reason. Um, all right. When you want a dentic governance [16:51] as a layer rather than as a project, come see us. We've been doing it for a little while. I did [16:55] want to make the point earlier that if you haven't been doing all this, I don't want to make you feel [16:58] bad. This is a newer technology and it's moving really, really fast. And we started at 02 not [17:04] that long ago. I'm not going to say how not long ago because I don't want you to know that, but we [17:08] had to get our hands around it too. Um, catalyst is an agent governance layer that we built to [17:13] rein in the uncovered agents, right? So, we can we can help you with that. Full agent observability, [17:17] spend tracking, knowledge preservation, that's a big one, that knowledge preservation, [17:21] policy enforcement, alerts, we can escalate to human in the loop when we need to. All of this [17:25] is something you can manage with the platform when you build on top of it. Now, that's enough about [17:30] Catalyst and this is my time for the day, but what I I do want to say real quick in passing [17:34] is I'm doing another talk tomorrow hopefully to at least 15 people, which would be really cool. [17:39] And and this is one as I was writing it last week and the reason why I'm going to take up a little [17:43] bit of air time here today about it that I think is important. And in case I don't get to see any [17:47] of you all at that one tomorrow, I'm going to say something right now about it. A lot of us that [17:50] are in the AI practitioning world right now are trapped in a bubble. And we love AI. And I truly [17:56] love AI. I think it it's going to be something that can be incredibly fantastic for all of [18:00] society. But it's not that easy. There's a lot of people that aren't being helped by AI right now. [18:06] It's a cultural problem. It's a society problem. a lot of people aren't given a choice about how [18:09] it's being used. And so I'm going to spend a little bit of my time during my talk tomorrow, [18:13] which I think the info is up there, talking about what we can probably do to come together [18:19] as a group of people who are the tip of the spear in this industry and help to make sure that AI is [18:24] helpful for everyone um and not just a few of us because I think that's critically important. [18:29] So that's my time everybody. I appreciate you being here. Really, I'm sorry that you didn't [18:33] really make it into the CIA or maybe you did. I don't know. Um but that's it. Thank you so much.