Give three frontier coding agents 200 frontend tasks with no guidance and they pass 55 percent. Add a curated pack of web platform knowledge and the same models pass 92 percent, with no retraining and no larger context window (Source: Irish, 2026).
That 37-point swing is the week's most useful architecture signal, because it points away from the thing most teams still tune. The constraint was not reasoning. It was whether the model had the current platform in front of it when it chose an API.
The Model Predicts the Modal Pattern, Not the Best One
The same failure mode shows up outside the browser. Researchers ran 28,125 completion prompts across seven LLMs and 145 API mappings from eight Python libraries, then measured how often a completion named a plausible API for the task: 10 to 30 percent (Source: Wang et al., 2024).
CodeLlama-7b topped the group at 30.6 percent, StarCoder2-3b bottomed out at 10.2 percent, and GPT-3.5, the largest model tested, landed at 15.2 percent (Source: Wang et al., 2024).
The paper names two causes. Deprecated calls sit in the training corpora, so models memorize the usage context, and at inference the model holds no posterior knowledge that the API was retired, so token probability wins (Source: Wang et al., 2024). The authors collected 9,022 functions using deprecated APIs against 19,103 using replacements, which explains why larger models improve without closing the gap (Source: Wang et al., 2024).
Ask an agent for a modal dialog and you often get a div, a focus trap, and 50 lines of JavaScript, the pattern that dominated the web for a decade. The platform ships <dialog> with showModal(), backdrop handling, and focus management, and models still do not default to it (Source: Irish, 2026).
You Cannot Prompt Your Way Out of a 1.7MB Gap
The obvious patch is a bigger system prompt, and it does not scale. The 1.0 pack is 200+ hand-written guides, roughly 1.7MB of text (Source: Irish, 2026). Loading that into every request spends context on guides a task will never need.
So the project ships an agent skill plus a retrieval CLI, with the search layer rebuilt from a 500MB native stack down to a roughly 7MB JavaScript package that answers in about 10ms on CPU (Source: Irish, 2026). Retrieval is cheap. Deciding when to retrieve is not.
That activation decision is where agent architectures quietly fail. Retrieval the model must remember to call never fires on weak models, and retrieval that fires every turn pays the token bill anyway. The team tuned the skill description to activate on frontend tasks and declined to cede activation to MCP dynamic tool search (Source: Irish, 2026).
The transferable rule: put the retrieval trigger in the harness contract, not in the model's judgment.
Deterministic Scanners and Agents Are Complements
Release 1.0 adds a modernization workflow that statically scans HTML, CSS, JavaScript, and package.json before the agent edits anything (Source: Chrome for Developers, 2026). The pattern classifier behind it trained on 30,000 GitHub repositories and reports 92 percent average precision on held-out repos (Source: Irish, 2026).
Scope is deliberately narrow: it flags legacy patterns it can detect cleanly and skips ambiguous ones to avoid false positives, covering about 20 percent of the guide set, with a stated target near 50 percent (Source: Irish, 2026). A deterministic detector is cheap and reproducible; a model reasoning over your code covers the tail at token cost, with answers that drift.
Designing a Knowledge Layer Into Your Own Stack
Version the knowledge artifact separately from the model. A guide set that updates with the platform, not with your vendor contract, stays current between model refreshes (Source: Chrome for Developers, 2026).
Put activation in the harness. Skill descriptions and explicit routing beat hoping the model calls a search tool it half remembers (Source: Irish, 2026).
Gate recommendations on your compatibility data. A primitive that ships in one engine is not automatically right; fallbacks exist for that case (Source: Chrome for Developers, 2026).
Keep a deterministic detector for patterns you can name. It is a stable, cheap floor under the model's variance (Source: Irish, 2026).
Ownership matters as much as content. A pack that steers every agent's frontend output is a standards artifact, so this one runs on shared seats: two Chrome, one Microsoft Edge, one Mozilla Firefox and MDN, and two independent expert seats, one held by Lea Verou (Source: modern-web-guidance, 2026).
Compatibility is delegated to Baseline, which calls a feature widely available only after it works across Safari, Chrome, Edge, and Firefox for at least 30 months (Source: Baseline, 2026). Guides carry that data, so an agent picks a modern primitive when your targets allow it and a fallback when they do not (Source: Chrome for Developers, 2026).
What Production Teams Actually Measured
Japan's note inc., with over 90 million monthly active users, ran a 180-run evaluation against its production codebase and moved modern web check pass rates from 61 percent to 94 percent. Across 431 merged pull requests it also lifted proper LCP image prioritization by 28 percent (Source: Chrome for Developers, 2026).
MetaMask installed the guidance as a review bot driven by a three-line prompt. In five weeks it reviewed roughly 480 pull requests, commented on 64, and engineers fixed 43 findings at a median of 15 minutes from comment to fix (Source: Chrome for Developers, 2026).
WeTV, Tencent's international streaming platform, put an agent in its CI/CD pipeline, where it now auto-resolves 90 percent of failed performance checks and cut debugging time 87 percent, from four hours to 30 minutes (Source: Chrome for Developers, 2026).
All three are vendor-published; treat them as directional. The durable part is the shared mechanism: none swapped the model, each changed what it could see.
FAQ
Q: Does this only apply to frontend work?
A: No. The study covers eight Python libraries and found the same pattern (Source: Wang et al., 2024). Fast-moving dependency surfaces inherit it.
Q: Will a newer model fix this on its own?
A: Partly, and not durably. The largest model in the study scored 15.2 percent plausibility while a smaller, Python-tuned model reached 30.6 percent (Source: Wang et al., 2024). Every model has a cutoff; the platform keeps shipping after it.
Q: How much context does retrieval actually cost?
A: Only the matched guides. The search layer installs at roughly 7MB and answers in about 10ms (Source: Irish, 2026).
Key Takeaway
The distance between mediocre and good agent output here was never reasoning capacity. It was 1.7MB of current, reviewed, versioned knowledge injected at the right moment, plus a compatibility gate so advice matched supported browsers. If a 37-point pass-rate jump came from an artifact your team could assemble, what is stopping you from building the equivalent for your own stack?
