Meta Slashes Video Player Crashes 80% Across Edits Suite

@Scale

Mobile decoders routinely misreport hardware capabilities, forcing developers to build defensive runtime fallbacks and autonomous agent loops to prevent creator churn.

AI-assisted summary of Preview Player Reliability and the Agentic Defense Loop on Edits | Bhumi Sabarwal & Ravi Shah — original source. Check the original for full context.

Source published: 2026-10-30

Key ideas

Meta's Edits app launched offensive engineering initiatives that cut unrecoverable playback failure rates by over 80% across platforms—dropping from >2.5% to ~0.38% on Android and 0.25% on iOS. (8:57)
Android playback stability suffered from OEM decoders misreporting capabilities, which caused 32% of Android decode failures before being addressed via codec leak fixes, smart software fallbacks, and bundled audio decoders. (5:36)
To prevent feature release regressions, the engineering team introduced Prepmate, an agentic defense loop running parallel sub-agents across diffs, error traces, and A/B experiments to deliver root-cause verdicts with confidence scores. (12:05)
Rather than fully automating incident response, the defensive system relies on a human in the loop to make final triage calls, such as rolling back an experiment or releasing a targeted fix forward. (13:13)

Detailed notes

## Background Published on YouTube by @Scale on October 30, 2026, this technical presentation features Meta engineers Bhumi Sabarwal and Ravi Shah discussing playback stability and automated triage on the mobile editing suite Edits. No additional show notes or description were provided with the release.

## TL;DR Bhumi Sabarwal and Ravi Shah outline their two-pronged strategy to fix preview player failures in Meta's Edits app, where unrecoverable playback errors initially affected over 2.5% of users. The team first launched an offensive engineering initiative addressing asymmetric platform bottlenecks, deploying recoverable projects, Android decoder fallbacks, and automated iOS AVFoundation composition resets to slash error rates by over 80%. To prevent feature velocity from introducing new regressions, they built Prepmate, an agentic defense system that deploys parallel sub-agents to analyze code diffs, error traces, and A/B experiments before escalating recommendations to a human engineer. Ultimately, they demonstrate that defensive architecture and zero-trust assumptions around low-level hardware decoders directly fuel creator trust and product growth.

## Atmosphere & style The presentation carries a crisp, pragmatic, and deeply technical engineering tone, moving seamlessly between macroeconomic creator stakes and low-level pipeline debugging. Bhumi and Ravi maintain a collaborative, disciplined rhythm, pairing systemic frustration over fragmented OEM decoders with calculated architectural triage. Watching the session feels like sitting in on a candid post-mortem and systems architecture review, blending hard-won mobile media debugging lessons with cutting-edge AI incident automation.

## Bulletin - [08:57](https://www.youtube.com/watch?v=btS_gNt5kEc&t=537s) Meta's Edits app launched offensive engineering initiatives that cut unrecoverable playback failure rates by over 80% across platforms—dropping from >2.5% to ~0.38% on Android and 0.25% on iOS. - [05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) Android playback stability suffered from OEM decoders misreporting capabilities, which caused 32% of Android decode failures before being addressed via codec leak fixes, smart software fallbacks, and bundled audio decoders. - [12:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=725s) To prevent feature release regressions, the engineering team introduced Prepmate, an agentic defense loop running parallel sub-agents across diffs, error traces, and A/B experiments to deliver root-cause verdicts with confidence scores. - [13:13](https://www.youtube.com/watch?v=btS_gNt5kEc&t=793s) Rather than fully automating incident response, the defensive system relies on a human in the loop to make final triage calls, such as rolling back an experiment or releasing a targeted fix forward.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=5s) Creator tooling reliability is directly tied to retention and livelihood, as 96% of creators earn under $100,000 annually and cannot tolerate broken creative progress. - [00:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=5s) Roughly 50% of content consumed on Instagram is driven by algorithmic recommendations rather than follower graphs, elevating the importance of raw production quality. - [02:43](https://www.youtube.com/watch?v=btS_gNt5kEc&t=163s) In H2 2025, telemetry showed that over 2.5% of Edits users experienced unrecoverable playback failures across file, codec, and renderer stages. - [05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) Failure root causes diverged heavily by platform: Android issues were spread across assets (43%), decoders (34%), and renderers (23%), whereas iOS failures were 93% player-level errors. - [05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) Edits introduced recoverable projects so missing or moved assets no longer hard-crash compositions, allowing users to relink or remove missing files directly in the timeline. - [05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) A core principle for Android media engineering is assuming lower layers lie about their capabilities, requiring smart runtime fallbacks and resource leak remediation. - [08:57](https://www.youtube.com/watch?v=btS_gNt5kEc&t=537s) On iOS, automatic re-instantiation of AVFoundation composition objects during Apple Media Services resets reduced player errors by 50% and boosted export success by 10%. - [12:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=725s) Offensive engineering work successfully lowered historical error baselines, but rapid feature velocity introduced continuous regression risks that manual triaging took over a week to resolve. - [13:13](https://www.youtube.com/watch?v=btS_gNt5kEc&t=793s) Prepmate orchestrates parallel AI sub-agents to analyze code diffs, error traces, and running experiments simultaneously when an alert triggers. - [13:13](https://www.youtube.com/watch?v=btS_gNt5kEc&t=793s) Agentic defense systems are bounded by telemetry fidelity, multi-experiment confounding interactions, and production-only regressions that evade pre-production detection. - [14:53](https://www.youtube.com/watch?v=btS_gNt5kEc&t=893s) Player reliability operates as an engagement growth metric, showing direct correlation between playback stability and increased user session activity. - [14:53](https://www.youtube.com/watch?v=btS_gNt5kEc&t=893s) AI acceleration shifts the engineering bottleneck from mechanical trace analysis to navigating semantic ambiguity and complex edge cases.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=5s) Why Creator Reliability Matters: The Stakes of Mobile Creation

The creator economy has transformed into a primary source of livelihood for millions around the world, backed by massive macroeconomic growth. Goldman Sachs projections indicate that the creation economy will near half a trillion dollars by 2027. At the same time, platforms like Instagram have shifted heavily toward algorithmic distribution: roughly 50% of the content consumed on Instagram is recommended algorithmically rather than driven by follower graphs. This shift places enormous weight on raw production quality. However, the economic reality for individual creators remains unforgiving, with 96% of creators earning less than $100,000 per year. Because profit margins are exceptionally thin and time is directly tied to income, creators cannot afford tooling issues.

To empower these creators, Meta launched Edits in 2025—a comprehensive, free mobile video editing suite. The app packs over 100 creative tools, supporting projects up to 10 minutes long with watermark-free exports that can be saved locally or shared directly across Instagram and Facebook. Core features include multi-clip timelines, filters, seamless transitions, and a standout AI cutout capability that quickly became a creator favorite, all underpinned by responsive playback and scrubbing.

Yet, creator loyalty in mobile software is fragile. A single playback freeze or unrecoverable error destroys creative progress and breaks user confidence. When a creation tool fails, users abandon the app and migrate to competing suites, directly hurting platform engagement and retention.

> “Reliability is user trust. The creator will come to the app once, but if they notice failures, they will not come again.” — Bhumi Sabarwal

> “You do not get a second chance to be the creator's creation app.” — Bhumi Sabarwal

- Creator economy market projection: Close to $500 billion ($0.5T) by 2027 (Goldman Sachs). - Instagram algorithmic mix: 50% of platform content is algorithmic rather than follower-driven. - Creator earnings: 96% of creators make under $100k annually. - Edits suite specifications: Launched in 2025; 100+ creative tools; supports up to 10-minute projects; watermark-free export across Meta apps and local storage.

---

[02:43](https://www.youtube.com/watch?v=btS_gNt5kEc&t=163s) Inside the Preview Player Architecture and Failure Modes

Bhumi introduces the session roadmap—spanning the internal preview player mechanics, the offensive engineering workstream, the defensive agentic loop, and upcoming roadmap items—before passing the stage to Ravi Shah to dissect the player under the hood.

The Edits preview player is constructed around three primary pipeline stages operating on an input media composition: 1. **Input Validation:** Inspects media compositions, checking asset validity, file availability, and platform read permissions. 2. **Demuxing & Decoding:** Extracts audio/video streams from containers and passes them to hardware or software decoders. 3. **Render & Composite:** Assembles decoded video frames, applies effects/cutouts/transitions, handles timing ranges, and pushes frames to the screen as a fluid preview.

```mermaid flowchart LR accTitle: Edits Preview Player Architecture accDescr: Pipeline showing media composition passing through validation, demux and decode, rendering, and final preview output.

subgraph Pipeline["Player Pipeline"] A["Media Composition"] --> B["Validate Input"] B --> C["Demux & Decode"] C --> D["Render & Composite"] end

D --> E["Preview Frames Display"] ```

When any of these pipeline stages break, the preview player freezes, leaving users stranded with generic "something went wrong" banners. In the second half of 2025 (H2 2025), telemetry revealed that over 2.5% of users experienced completely unrecoverable playback failures when playing videos in the Edits app.

These errors broadly map into three categories: - **File / Asset Errors:** Input validation failures caused by missing media assets, corrupted source files, or file-system permission rejections. - **Codec Errors:** Demuxing and decoding failures triggered by formats incompatible with the target device hardware or decoders misreporting capabilities. - **Renderer Errors:** Composition crashes caused by invalid trim ranges, out-of-bounds frame math, or `NaN` (Not-a-Number) time ranges.

---

[05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) Offense Engineering: Diagnosing Platform Asymmetries and Android Codec Pitfalls

To aggressively lower the >2.5% failure rate, the engineering team launched an "offense" workstream targeting root causes across platforms. Telemetry uncovered striking divergences between Android and iOS failure distributions: - On **Android**, failures were broadly distributed: 43% were asset-related (missing files, revoked access permissions), 34% were player/decoder-related, and 23% were renderer-related. - On **iOS**, the profile was heavily skewed: 93% of user failures were player-level errors, while asset errors accounted for only 5%, and renderer errors were virtually negligible at roughly 1%.

The team implemented distinct fixes for each platform's architectural bottlenecks.

#### Fix 1: Recoverable Projects (Asset Handling) Instead of hard-crashing an entire project when a source video was deleted from device storage or moved, Edits introduced "recoverable projects." The app exposes in-line UI enabling creators to directly salvage their composition by either deleting the unresolvable asset reference or relinking/replacing it with a compatible file, preserving the rest of the edit timeline.

#### Fix 2: Android Codec Defenses Android OEMs frequently misreport media decoding capabilities. As Ravi explains: “The device lies about what it can decode.” Android devices presented fragmented decoder configurations, strict concurrency limits, and OEM drivers claiming support for video formats they could not decode in practice, accounting for 32% of Android user decode failures. The team deployed three specific interventions: 1. **Codec Leak Remediation:** Plugged resource leaks where media codecs failed to release hardware instances, yielding a 40% reduction in codec errors. 2. **Smart Codec Fallback:** Dynamically switches between hardware and software decoders based on real-time runtime state, device tier, and format compatibility, driving an additional 22% error reduction. 3. **Bundled Decoders:** Bundled dedicated audio decoders directly alongside video decoding pipelines, cutting decoding failures by another 13%.

---

[08:57](https://www.youtube.com/watch?v=btS_gNt5kEc&t=537s) Fix 3 on iOS: AVFoundation Resets and Massive Platform Gains

On iOS, the dominant failure was the notorious `AVPlayer` / Apple Media Services reset error, where underlying media daemons crash or disconnect, corrupting active `AVFoundation` objects. Rather than failing the session, Edits implemented an automated system that reconstructs the entire media composition from scratch on the fly. When a Media Services reset occurs, the player re-instantiates clean `AVFoundation` objects (`AVAsset`, `AVPlayer`, composition tracks) and feeds a fresh copy of the asset back into the pipeline without user-facing disruption.

This iOS reset mechanism produced outsized wins: - A 50% direct reduction in iOS player errors. - A 2% latency/performance improvement across the timeline playback. - A 10% increase in export success rates. Transcoding pipelines—which perform heavy background rendering during exports—frequently crashed when inheriting degraded or corrupted asset handles; feeding fresh, recreated assets gave export workers significantly more headroom to finish encoding reliably.

Across both mobile platforms, the offensive campaign drove error rates down by more than 80%: - **Android Error Rate:** Reduced by 82%, dropping from >2.5% down to approximately 0.38%. - **iOS Error Rate:** Reduced by 90%, falling from 2.5% down to 0.25%.

> “Offense definitely raises the bar but we also need defense to continuously hold the bar.” — Bhumi Sabarwal

---

[12:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=725s) Transitioning to Defense: Introducing Prepmate

While the offense workstream successfully knocked down historic baseline errors, rapid feature development introduced a new operational crisis. Rapid product adoption and feature velocity meant an influx of code diffs into every release train. Production regressions spiked, causing fatal SEVs (site events). Manually triaging, identifying culprit commits, and resolving these regressions routinely took more than a week per incident.

To automate incident mitigation and sustain reliability gains, the team engineered **Prepmate**, an agentic defense loop designed around three core operational stages: 1. **Real-Time Detection:** Continuous monitors across pre-production and production channels that fire automated alerts whenever anomaly thresholds are breached. 2. **Parallel Agentic Fan-Out:** Upon alert triggering, the system dispatches three autonomous AI agents simultaneously, powered by custom AI recipes: - *Error Analysis Agent:* Deconstructs stack traces, exception logs, and failure signatures. - *Code Diff Agent:* Analyzes recent repository commits and landing diffs to pinpoint potentially causative changes. - *Experiment Analysis Agent:* Evaluates running feature gates and A/B experiments to isolate regressing test groups. 3. **Synthesis & Mitigation:** Collects findings across all three parallel agents to deliver rapid, actionable diagnostics.

[13:13](https://www.youtube.com/watch?v=btS_gNt5kEc&t=793s) The Agentic Defense Loop: Workflow and Limitations

Bhumi Sabarwal walks through how the agentic defense loop operates when an actual issue occurs in the system. The workflow begins when an alert fires in pre-production. In response, an on-call agent automatically triggers three specialized sub-agents running in parallel: one dedicated to diff search, another performing error analysis, and a third conducting experiment analysis. Once these parallel sub-agents return their findings, the central orchestrator synthesizes and cross-references the data to isolate the root cause. Crucially, the system does not act entirely autonomously without oversight; rather, it produces a verdict accompanied by a confidence score, bringing an engineer into the loop at the very end to decide whether to deallocate the offending experiment or push forward a targeted fix. As Bhumi Sabarwal explains, “Once there is a verdict, it will be sent across with a confidence number and that is the first time the human is involved to take the decision forward.”

```mermaid flowchart TD accTitle: Pre-Production Agentic Defense Loop accDescr: An alert in pre-production triggers an on-call agent which runs three parallel sub-agents before an orchestrator synthesizes the results for human decision-making.

subgraph Detection["Detection Phase"] A["Pre-Production Alert Fires"] --> B["On-Call Agent Triggered"] end

subgraph ParallelAgents["Parallel Sub-Agents"] B --> C1["Diff Search Agent"] B --> C2["Error Analysis Agent"] B --> C3["Experiment Analysis Agent"] end

subgraph Synthesis["Synthesis & Resolution"] C1 --> D["Orchestrator Synthesizes & Cross-References"] C2 --> D C3 --> D D --> E["Verdict with Confidence Number"] E --> F{"Human in the Loop"} F --> G["Deallocate Experiment"] F --> H["Pick Up Fix Forward"] end ```

Despite the efficiency of this multi-agent setup, Bhumi outlines several clear boundaries and operational limitations that prevent full automation. First, the agents are inherently limited by telemetry quality; uninformative stack traces that omit granular error details leave the agents without sufficient signal. Second, there are insidious "slow boiling" regressions that manage to bypass pre-production testing completely and only manifest under the volume and variety of real-world production traffic. Third, complex interactions spanning multiple concurrent experiments can confound the system, making it nearly impossible for an agent to disentangle which specific experiment triggered the regression. Finally, the tool is only as effective as its configuration: if the diagnostic lens and agent specialization are not tailored precisely to the domain, the system will fail to uncover the regression.

- **Trigger:** Pre-production alerts initiate the automated on-call triage loop. - **Parallel processing:** Three sub-agents simultaneously investigate code diffs, error traces, and active experiments. - **Human-in-the-loop validation:** Human intervention is reserved for final decision-making, informed by synthesized root-cause verdicts and confidence ratings. - **Known limitations:** Gaps in trace details, multi-experiment interaction confounds, production-only regressions, and improperly tuned agent lenses.

---

[14:53](https://www.youtube.com/watch?v=btS_gNt5kEc&t=893s) Core Takeaways and Creator Impact

Ravi Shah takes over to summarize the broader architectural and organizational lessons learned from hardening preview player reliability and building the agentic loop on Edits. His first major takeaway is that reliability functions directly as a growth metric; the team observed that as performance and playback stability were systematically resolved, user engagement climbed in tandem. His second principle stems directly from their experiences debugging hardware and lower-level media stacks: always assume that the lower system layers are being untruthful about their operational state. As Ravi Shah puts it, “As we mentioned earlier in our talk that decoder they might claim something but does not work as the same way. So we always assume the layer below is always lying.”

His third takeaway addresses the role of machine learning in engineering operations: AI acceleration does not eliminate bottlenecks, but rather shifts where they sit. While automated agents eliminate massive amounts of manual diffing and trace analysis that engineers struggle to execute quickly at scale, the ultimate bottleneck shifts to handling edge cases and semantic ambiguity, preserving an indispensable role for human engineers. Ravi concludes by highlighting the direct response from content creators using Edits. Creators have felt a clear transformation in day-to-day usability, noting marked gains in speed and preview reliability. This shift has altered product sentiment, elevating Edits into a dependable, mature creative tool where frequent complaints about broken or unplayable videos have largely disappeared.

- **Reliability drives growth:** Resolving playback performance roadblocks led to measurable, sustained increases in user engagement. - **Zero trust for lower layers:** Decoders and platform-level APIs frequently report incorrect states, requiring defensive validation higher up the stack. - **AI shifts the bottleneck:** Automation absorbs mechanical analysis, while humans remain essential for resolving high-ambiguity edge cases. - **Creator sentiment turnaround:** A steep decline in core playback failure reports ("my video is not playing" or "it is broken") has repositioned Edits as an advanced, reliable editing platform.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=btS_gNt5kEc&t=5s) You do not get a second chance to be the creator's creation app. — Bhumi Sabarwal

> [05:36](https://www.youtube.com/watch?v=btS_gNt5kEc&t=336s) The device lies about what it can decode. — Ravi Shah

## Mentioned Bhumi Sabarwal, Ravi Shah, Edits, Meta, Instagram, Facebook, Goldman Sachs, Prepmate, AVFoundation, AVPlayer, Apple Media Services, Android

[Source](https://www.youtube.com/watch?v=btS_gNt5kEc)

中文简报

Meta将Edits套件视频播放器崩溃率削减80% 移动端解码器经常误报硬件性能，迫使开发者构建防御性运行时回退机制与自主智能体闭环，以防止创作者流失。 Bhumi Sabarwal与Ravi Shah概述了修复Meta旗下Edits应用中预览播放器故障的双管齐下策略；此前，不可恢复的播放错误曾影响超过2.5%的用户。该团队首先发起了一项攻坚工程，针对不对称的平台瓶颈，部署了可恢复工程、Android解码器回退机制以及自动化iOS AVFoundation合成重置，从而将错误率削减了80%以上。为了避免功能迭代速度引发新的回归，他们构建了Prepmate智能体防御系统，该系统部署并行子智能体来分析代码差异、错误堆栈和A/B实验，随后向人工工程师提交建议。最终，他们证明了针对底层硬件解码器的防御性架构与零信任假设能直接增强创作者信任并推动产品增长。 Meta的Edits应用发起了攻坚工程，在各平台将不可恢复的播放故障率削减了80%以上——在Android端从>2.5%降至约0.38%，在iOS端降至0.25%。 (8:57) Android端播放稳定性曾因OEM解码器误报性能而受损，这导致了Android端32%的解码故障，随后团队通过编解码器泄漏修复、智能软件回退机制和捆绑式音频解码器解决了该问题。 (5:36) 为防止功能发布引发回归，工程团队引入了Prepmate智能体防御闭环，通过跨代码差异、错误堆栈和A/B实验运行并行子智能体，提供附带置信度评分的根本原因判定。 (12:05) 该防御系统并未实现完全自动化的事件响应，而是依靠人工介入进行最终的分流决策，例如回滚实验或发布针对性的前向修复。 (13:13)

Reading summary

To protect mobile creator retention, Meta's Edits team reduced preview player playback failure rates by over 80% through targeted platform-level fixes, subsequently implementing an agentic defense loop called Prepmate to catch continuous regressions. Creator Economics and the Stakes of Mobile Tooling (0:05) The creator economy is projected by Goldman Sachs to approach $500 billion by 2027, with platforms like Instagram shifting roughly 50% of consumed content to algorithmic recommendations rather than follower graphs. Despite this macroeconomic scale, 96% of creators earn less than $100,000 annually, operating on tight margins where editing tool failures directly threaten income. Preview Player Architecture and Asymmetric Platform Failures (2:43) The Edits preview player operates across three main stages: input validation, demuxing and decoding, and rendering and compositing. In the second half of 2025, telemetry indicated that over 2.5% of users experienced unrecoverable playback failures from missing assets, decoder incompatibilities, or renderer time-range errors. Diagnostics uncovered striking platform asymmetries. On Android, user failures were distributed across asset handling (43%), player decoding (34%), and rendering (23%). In contrast, iOS failures were heavily concentrated, with 93% stemming from player-level errors, 5% from assets, and only 1% from rendering. Platform Engineering Interventions on Android and iOS (5:36) Because Android decoders frequently misreport their actual capabilities, the team applied three targeted interventions: plugging codec resource leaks, dynamically falling back between hardware and software decoders, and bundling dedicated audio decoders. These adjustments reduced Android codec errors by 40%, 22%, and 13% respectively. Additionally, recoverable project workflows were introduced so missing or moved files no longer crashed entire compositions. On iOS, where the underlying Apple Media Services daemon frequently crashed and corrupted active AVFoundation objects, Edits introduced an automated pipeline that reconstructs composition objects from scratch on the fly. This mechanism cut iOS player errors by 50%, increased export success rates by 10%, and improved timeline playback performance by 2%. Overall, Android error rates dropped by 82% to approximately 0.38%, while iOS error rates fell by 90% to 0.25%. Automating Regression Triage With Agentic Defense Loops (12:05) To prevent rapid code shipping from reintroducing fatal site events that historically took over a week to triage, the team developed Prepmate. When a pre-production alert triggers, an on-call orchestrator fans out to three specialized sub-agents running in parallel: an Error Analysis Agent, a Code Diff Agent, and an Experiment Analysis Agent. The orchestrator synthesizes the findings to output a root-cause verdict alongside a confidence score. Human engineers are intentionally kept in the loop at the final stage to make the decision to deallocate an experiment or push a targeted fix forward, shifting engineering effort away from manual diffing toward higher-ambiguity edge cases. The automated defense loop remains bounded by operational constraints: it struggles with uninformative stack traces, regressions that bypass pre-production and appear only under production traffic, and compound interactions across concurrent experiments. Engineering leadership emphasizes that developers must assume underlying hardware and platform layers are untruthful about their operational status.

精炼阅读摘要

Meta 旗下视频剪辑应用 Edits 团队通过底层架构的主动防御优化与基于 AI Agent 的被动防御闭环，将移动端播放失败率降低了 80% 以上，并确立了“可靠性即增长动力”的工程实践原则。创作者经济与预览播放器的核心挑战 (0:05) 创作者经济正向近半万亿美元规模扩张，高盛预测其到2027年将接近5000亿美元。在 Instagram 等平台上，约50%的内容分发来自算法推荐而非粉丝关注，这使得内容制作质量变得至关重要。然而，96%的创作者年收入不足10万美元，利润空间微薄且时间紧迫，导致创作者对制作工具的容错率极低。Meta 于2025年推出包含100多种工具、支持最长10分钟无水印导出的移动剪辑软件 Edits，但移动端工具的单次崩溃便可能导致用户彻底流失。Bhumi Sabarwal 强调，可靠性即用户信任，工具一旦失败将失去成为创作者核心应用的机会。 Edits 的预览播放器包含三级核心流水线：输入校验（检查资产有效性与读取权限）、解复用与解码（提取音视频流送入软硬件解码器）以及渲染合成（处理特效、抠图与转场）。遥测数据显示，2025年下半年，超过2.5%的用户遭遇完全无法恢复的播放失败。这类故障主要由三方面引发：丢失媒体或权限被拒导致的资产错误、解码器能力不符或格式不兼容导致的编解码器错误，以及由非法剪辑区间或 NaN 时间范围引发的渲染器崩溃。进攻性架构改造与跨平台差异化修复 (5:36) 遥测数据揭示了 Android 与 iOS 平台间显著的故障分布差异：Android 端错误较为分散，43%为资产问题，34%为解码器问题，23%为渲染器问题；而 iOS 端 93% 的故障集中在播放器与解码层，资产错误仅占5%，渲染错误仅约1%。针对资产丢失导致整个工程直接崩溃的情况，团队开发了“可恢复项目”功能，允许创作者通过删除失效引用或替换重连可用文件来挽救编辑工程，避免项目彻底损坏。在 Android 平台上，OEM 驱动常虚报解码能力，硬件解码碎片化与并发限制占解码故障的32%。Ravi Shah 指出“设备会对其解码能力说谎”。团队针对性实施了三项措施：修补媒体编解码器的硬件实例泄漏，减少了40%的编解码错误；根据运行时状态与设备档次引入软硬件动态回退，进一步降低22%错误；将专用音频解码器与视频管线绑定，再削减13%故障。在 iOS 端，针对底层媒体服务崩溃导致 AVFoundation 对象损坏的问题，团队构建了动态重建机制，在服务重置时自动无感实例化全新的 AVAsset 与播放对象。此举使 iOS 播放器错误减少50%，时间线播放延迟改善2%，重度依赖底层资产句柄的导出转码成功率提升10%。总体上，Android 端故障率下降82%至约0.38%，iOS 端故障率下降90%至0.25%。基于 Prepmate 的 Agentic 防御闭环 (12:05) 进攻性工程虽然大幅压低了历史基线故障率，但随着功能快速迭代与大量代码合并，生产回归问题频发，常规人工排查单一故障往往耗时一周以上。为此，团队构建了自动化防御系统 Prepmate。该系统由预生产与生产监控的实时告警触发，一旦指标突破异常阈值，中央调度器会同时并行唤起三个专门的 AI Agent：错误分析 Agent 负责解析调用栈与异常签名；代码差异 Agent 审查近期代码提交以锁定可疑变更；实验分析 Agent 评估进行中的 A/B 实验分桶以排查回归来源。三大子 Agent 返回诊断结果后，编排器会汇总交叉验证并输出带有置信度分数的判定报告。Bhumi Sabarwal 明确指出，系统并非全自动处理，而是在最终环节交由值班工程师把关，由人类决定下线实验还是推进定向修复。同时，该 Agent 体系存在明确的运行边界与局限：缺少详细信息的堆栈调用会导致分析失效；绕过预生产并在真实生产流量下才爆发的慢性回归难以提早拦截；多个并行实验间的复杂交互会形成干扰；此外，若诊断侧重点与专业化配置未精确适配业务领域，系统也会错失回归归因。工程认知重塑与创作者体验转变 (14:53) Edits 团队在经历播放器可靠性改造后总结出三项核心工程理念。首先，可靠性直接等同于增长指标，播放卡顿与故障的清除带来了用户参与度的可衡量增长。其次，针对底层硬件与系统服务必须采取“零信任”原则，默认下层系统可能虚报自身状态，必须在应用层建立防御校验机制。再者，AI 的引入并未消除工程瓶颈，而是转移了瓶颈——自动化接管了繁重的机械化差异审查与日志比对，但复杂边界情况与语义模糊判定仍高度依赖人类工程师。这一系列攻防体系的落地显著扭转了创作者的负面反馈。此前频繁出现的“视频无法播放”或“工程损坏”投诉大幅减少，预览速度与稳定性的提升让创作者切实感受到了产品性能的质变，推动 Edits 转型为稳定可靠的专业级创作平台。该工程方案虽通过进攻性重构与 Agentic 防御闭环消除了八成以上的核心播放失败，但在调用栈信息缺失、多重实验并发干扰以及仅在特定高流量下暴露的慢性回归场景中，自动化归因仍面临盲区，人机协同与底层防御性校验依然是维持系统稳定的必要防线。

Source quotes

[0:05] Bhumi Sabarwal: You do not get a second chance to be the creator's creation app. [5:36] Ravi Shah: The device lies about what it can decode.

Original transcript

[0:05] Let's start with a very basic question. [0:08] Why creation matters? Creation is how a [0:12] number of people are earning a living [0:14] today. According to a research from [0:16] Goldman Sachs, the creation economy will [0:19] be close to half a trillion dollars by [0:22] 2027. [0:23] Also 50% of the content on Instagram is [0:28] algorithmic and not followerdriven [0:31] which goes to say that the content and [0:34] its quality is really important. [0:37] 96% of the creators earn under 100k a [0:41] year. The margins are really thin and [0:44] the reliability of the creation app is [0:47] very important. [0:51] So what is Meta doing for creators? [0:54] Meta has built edits which is a powerful [0:57] creation suite which is available on [1:00] phones and for free. Let me take you [1:03] through a demo. As you can see the user [1:07] is able to add multiple clips, [1:09] transitions, filters and even cutouts [1:13] which is a very new feature that our [1:16] creators love. [1:17] Also notice the smooth playback and seek [1:20] experience in the app. [1:23] Edits was launched in 2025. [1:26] There are 100 plus creative tools in the [1:29] suite. 10 minutes of projects with [1:32] watermark free exports that can be [1:35] shared across Facebook, Instagram, and [1:38] can also be downloaded locally. [1:42] Creators love edits. [1:44] In their own words, they find it easier [1:47] to use. They switch from other apps. [1:50] Cutouts, which is the feature that we [1:52] just saw, is highly loved. And they find [1:55] the app reliable and working all the [1:57] time. [2:00] Reliability is user trust. The creator [2:04] will come to the app once, but if they [2:06] notice failures, they will not come [2:08] again. [2:10] you do not get a second chance to be the [2:12] creator's creation app. [2:18] So this is what happens when there is a [2:20] failure. You just see a error message [2:22] and the playback is just stuck. [2:26] In their own words, creators have [2:29] reported lost work, broken trust, and [2:33] they often have to go to other apps to [2:36] complete their creation, which is [2:38] detrimental to the engagement of the [2:40] app. [2:43] This brings us to our topic for today, [2:47] preview player reliability and agentic [2:49] defense loop on edits. [2:52] Hi, I'm Roomie and today along with [2:55] Rabbi, we will be taking you through the [2:57] preview player on edits. We will take a [3:00] sneak peek into the player, our offense [3:03] work stream, our defense agentic loop, [3:07] what is coming up next and some [3:10] takeaways. [3:12] I will now hand it over to Ravi to take [3:14] us through the player. Hey, [3:16] >> thanks Kumi. Hi all, my name is Ravi and [3:19] I will dive into what is inside the [3:20] player. [3:23] So what is inside the player? It made of [3:26] three major components. You have a input [3:29] as a media composition. [3:32] That first component it does the [3:34] validate the input for you. Second one [3:38] does gimmax and decoding and the third [3:41] one does render and composite your [3:43] frames. So those are the major three [3:45] components that we have and everything [3:47] binded together we will have one frame [3:50] or series of frame that you will see on [3:52] on your screen which we call as a [3:54] preview. So those are the major [3:57] components inside our player. Now what [4:01] can goes wrong in the player right? So [4:03] the first one is while validating the [4:06] input we may see file errors missing [4:09] assets or file is invalid or you know [4:13] like we cannot continue with this file [4:16] or because of permission issues. So all [4:19] these are validate input around file [4:21] errors. Second one is we are doing demox [4:24] and decode. we might get into the codec [4:27] errors [4:29] is seem like is incompatible for the [4:31] codec or it is incompatible this file [4:34] for this device. So we might run into [4:37] the codec errors and the third one is [4:39] about renderer errors. So you might see [4:43] we have a trim range error or we have a [4:46] not a number kind of time range. So all [4:50] these are the renderer errors. So those [4:53] are the major three types of error. [4:57] what we are doing with this error or [4:59] like what is impacting creators for this [5:02] error. So in the H2 2025 [5:07] we have found out that more than 2.5% of [5:11] users are basically getting into this [5:15] error which is unreoverable when you [5:18] play any video in edits app. Right? So [5:20] if you look at on the image on the right [5:22] side user click on a play button and you [5:26] can see something went wrong like so you [5:29] will not see any frames are getting [5:30] updated it looks more frozen video and [5:33] nothing nothing works there. [5:36] Okay. So what we are doing on the [5:39] offense side to help those creators to [5:42] do more reliable editing inside the edit [5:45] app. [5:47] So we went back and then we we thought [5:50] about like what where are the failures [5:53] coming from right so you can see on [5:55] Android and iOS there are major three [5:58] categories of failure set failure player [6:01] failure and renderer on Android we have [6:06] around 43% of failures which is related [6:08] to asset like missing file no permission [6:11] to play on iOS we have around only 5% of [6:16] user hitting those errors. But the [6:19] surprise that we got on the iOS is that [6:22] around 93% of users hitting the player [6:26] related errors while on Android is like [6:29] one/ird of users were hitting around the [6:31] player related errors. What stands out [6:35] from this one is on iOS the render error [6:39] is really minimal like 1% of users were [6:42] hitting while on I while on Android we [6:45] have around 23% of users were hitting [6:48] those renderer errors. So what we have [6:52] done to fix those failure I will walk [6:55] you through through two failures and [6:58] we'll show how we fix it. [7:01] The first fix that we have done through [7:03] the recoverable projects. So recoverable [7:06] projects that we are giving a creators [7:10] control to fix their composition. when [7:13] we have a set errors. So set error as I [7:16] mentioned earlier that could be file [7:18] missing, file cannot open, permission [7:21] issue, you provide a user or creators to [7:25] fix their composition either by deleting [7:27] the file or by replacing the file which [7:30] is incompatible to play. So user fix [7:34] those file by replacing and it will [7:38] continuing your playback inside the edit [7:41] s and they creators continue reliably [7:43] editing their their videos. [7:47] On the fix number two which is more [7:50] focus around Android the device lies [7:53] about what it can decode. So if you [7:56] remember I mentioned inside the player [7:58] that there is a component where we demox [8:01] and decode your video. We have seen that [8:04] on Android specifically we have seen [8:06] like fragmented decoders are concurrency [8:09] limits format they claim OEM but cannot [8:12] play and around 32% user were falling in [8:16] this category that cannot decode for the [8:19] Android. So we have made three fixes on [8:22] Android. First one we made a fix about [8:25] codec leaks that improve the error by [8:28] 40%. [8:29] Second fix that we made about smart [8:31] codec fallback. In the smart codec [8:34] fallback either you use hardware codec [8:37] as a fallback or software codec as a [8:39] fallback depending on the device and [8:41] condition that improves this error by [8:44] 22%. And the third one that we have used [8:47] the bundle decoders. We have bundle A [8:50] decoder for audio with the video [8:52] together that reduce the the decode [8:55] failures by 13%. [8:57] Fix number three which is more focused [9:00] toward the iOS. We have fixed the Apple [9:03] media service reset errors by recreating [9:06] everything your media composition from [9:08] the scratch. So if you remember from my [9:10] inside the player diagram the input that [9:12] we get is a media composition. While we [9:15] hit Apple media service reset error, we [9:17] need to reload all the assets that we [9:20] have created to make sure that it does [9:22] have a phrase copy of the asset. So for [9:25] an example any AV foundation objects [9:28] that we create like a asset, a player or [9:31] anything that we create, we have to [9:32] recreate everything and then we fix [9:36] those errors on the fly. The user will [9:38] not feel any disruption while editing [9:41] their video. This fix comes with some [9:45] wins. Major win that we got around 50% [9:48] of error reduction for the iOS failures. [9:52] Second is about we have seen 2% [9:55] improvement on our timeline which is [9:57] huge. And the third one which is more [10:00] interesting we got around 10% [10:02] improvement on export success rate. And [10:06] this is more surprising because we have [10:09] thought about that it is coming as a [10:12] more time or more room for export [10:16] operations that are doing heavy lifting [10:20] for like transcoding. They get more room [10:23] with the phrase asset that we have feed [10:26] into the system and then they will go [10:30] and use those asset as a transcode [10:33] transcode input and they continuing [10:36] encoding and then generate the file out [10:38] of it and 10% improvement was a huge for [10:41] the export. It improves user exporting [10:44] the video as well not just the timeline [10:46] or the preview player. [10:48] Okay. So where we are with all those [10:50] fixes right so as I mentioned earlier we [10:52] started on both platform more than 2.5% [10:55] users are feeling the issues on Android [10:58] we have fixed around 82% of those error [11:03] and currently we are around 38% users [11:07] are failing or have a you know like bad [11:09] user experience on iOS we have improved [11:12] by 90% going from 2.5% to like 0.25%. [11:17] There's a huge improvement across both [11:20] platforms on Android and iOS. They are [11:23] seeing like more than 80% of improvement [11:26] on both platform. [11:28] Now the creators can reliably and [11:32] smoother editing on the edit tab without [11:36] worrying about you know like losing [11:37] their [11:39] losing their composition or losing their [11:41] media or their work. Offense raises the [11:44] bar. Now I will hand over to Bumi. Civil [11:47] talks about defense. [11:49] >> Thanks Ravi. So we just saw our offense [11:52] work stream that helped us resolve all [11:55] the top errors. [11:57] Offense definitely raises the bar but we [12:00] also need defense to continuously hold [12:03] the bar. [12:05] Why offense alone stopped being enough? [12:08] With a huge adoption and new features, [12:11] we also saw increase in production [12:14] regressions. [12:17] A lot more diffs started to land in each [12:20] release. There were multiple fatal site [12:22] events and it was taking us more than a [12:25] week to resolve these regressions [12:28] manually. [12:30] Hence, we built Premate which is our [12:32] agentic defense loop. [12:35] Prepmate consists of three stages. The [12:38] first is the detection where we have [12:41] realtime production and pre-production [12:43] alerts that fire at certain thresholds. [12:47] Next, once the alert fires, it will fan [12:50] out three agents in parallel to do the [12:53] error analysis, to look up the code [12:56] diffs that could be related to the [12:58] errors, and to also do experiment [13:01] analysis to find out any regressing [13:03] experiments. [13:05] These agents are powered by custommade [13:08] AI recipes. [13:10] Once all the three agents return, it [13:13] will synthesize and act to cross [13:16] reference all the findings and find a [13:18] root cause. [13:21] Finally, we will either deallocate the [13:23] experiment or pick up a fix forward. [13:29] Let us walk through an actual issue. [13:33] So let's say an alert fired in [13:36] pre-production. [13:38] At that time a on call agent would fire [13:42] that would in turn fired three sub aents [13:45] in parallel to do diff search error [13:48] analysis and experiment analysis. [13:52] Once all the agent results are back, the [13:54] orchestrator will cross reference all [13:57] their findings to figure out the root [13:59] cause. [14:01] Once there is a verdict, it will be sent [14:03] across with a confidence number and that [14:06] is the first time the human is involved [14:08] to take the decision forward. [14:12] There are always limits to these agents. [14:16] There could be uninformative traces that [14:19] lack the error details. [14:22] There could be slow boiling regressions [14:24] that do not show up in pre-production [14:27] but only show up in production. [14:30] There could be multi-experiment [14:31] interactions, but the agent cannot [14:34] figure out the regression from a [14:36] particular experiment [14:38] and the lens is only helping if it is [14:41] applied correctly. If the agent is not [14:44] customized correctly, it will not be [14:46] able to figure out the regression. [14:49] I will now let Rabi take us through [14:52] what's coming up next. [14:53] >> Thanks, Bumi. In closing, I will take [14:56] you through some takeaways. Number one, [15:00] reliability is a growth metric. We have [15:03] seen while we are fixing all the [15:05] improvements on the performance side, [15:07] the user engagement keep growing over [15:10] the period of time. Second one is assume [15:14] the layer below is always lying. As we [15:16] mentioned earlier in our talk that [15:18] decoder they might claim something but [15:21] does not work as the same way. So we [15:23] always assume the layer below is always [15:26] lying. [15:27] Third one is about AI acceleration that [15:30] moves the bottleneck. So AI always helps [15:34] around the bottleneck. It moves the [15:35] bottleneck. The human cannot do a lot of [15:38] stuff by themsel. AI helps to move the [15:41] bottlenecks but human will come into the [15:44] picture to handle the ambiguity. So [15:47] human will always be in the loop to [15:49] handle any ambiguity from the agent. [15:53] Now when we made all the fixes, the [15:56] content creator have noticed all the [16:00] work that we have done so far. The first [16:02] one is they see noticeable improvements [16:06] in both speed and reliability. [16:10] Second one, the perception has shifted. [16:13] They think that edits is more mature app [16:15] right now. [16:17] Third one, now we are seeing fewer [16:19] complaints about reliability like my [16:21] video is not playing or you know like [16:23] it's is broken. [16:26] And the fourth one is now that it's [16:28] ahead of reliability. [16:31] Thank you everyone for tuning in. I will [16:33] see you next time.