self.md radar — 2026-05-10
Agents are being pushed into the messy middle now: long edits, browser runs, and rendered artifacts, not just answers.
The useful signals today all pointed at the same operating problem. DELEGATE-52 measured document drift under delegation; HyperFrames and the HTML artifact push made agent output easier to inspect; BrowserTrace gave browser agents a local flight recorder.
1. delegation has a corruption rate
sources:
what happened: DELEGATE-52 puts 52 long document-editing workflows in front of 19 LLMs, across domains from coding to crystallography to music notation. The paper reports that even frontier models corrupt an average of 25% of document content by the end of long delegated workflows. Agentic tool use does not fix it; the worse cases come from larger documents, longer interactions, and distractor files.
why this matters: A delegate that fails loudly wastes a run. A delegate that makes sparse, severe edits inside a document and hands it back clean changes the review job: you need drift checks around the work, not a polite final summary.
2. HTML is turning into agent output glue
sources:
what happened: HyperFrames is an open-source video renderer built around one blunt promise: write HTML, render video, and let agents do the composition work. Its repo ships skills or plugin surfaces for Claude Code, Cursor, Gemini CLI, and Codex, including GSAP timelines, Tailwind v4 browser-runtime styles, and first-party animation adapters. The HTML artifact demo page points the same way with annotated PRs, module maps, design token sheets, clickable flows, slide decks, incident timelines, and prompt tuners as single-file HTML.
why this matters: Markdown is still good for notes. Once the agent is producing video, review surfaces, UI flows, or design evidence, the useful object is closer to a tiny web app: visible, inspectable, screenshot-able, and easy to feed back into the next run.
3. browser agents got a flight recorder
sources:
what happened: BrowserTrace records each browser-agent step as a timeline with screenshot, URL, action, model input, model output, status, and error. It is built for Browser Use, Stagehand, Skyvern, Playwright plus LLM scripts, and custom computer-use agents. The CLI can inspect a run from the terminal or export a public-safe HTML report that omits model I/O, screenshots, and URLs.
why this matters: For browser agents, “it clicked the wrong thing” is not a bug report. The evidence disappears with the page state unless the run keeps a trace of what the model saw, what it was told, and which action moved the system off the rails.
supporting links
- ContextChef — context compiler for TypeScript and JavaScript agents with history compression, tool pruning, memory, VFS offloading, and provider adapters.
- Draft — Claude Code, Codex, and Cursor plugin that keeps product context through an append-only log and persistent index.
- Snyk with Claude Code — security scanning moving into the coding-agent loop instead of waiting for a finished PR.
- Remind — local Mac workflow that schedules Claude Code through Apple Reminders and writes the result back into the reminder.
left on the table
- oh-my-pi had the biggest repo gravity, but the seen ledger already has it on 2026-03-08 and 2026-04-24; a push today was not enough reason to rerun the same terminal-agent story.
- Firefox using Claude Mythos for bug hunting stayed out because the recent radar already covered both the Firefox hardening surface and the METR/Mythos measurement wave.
- ContextChef and Draft were useful support, but making either one a main signal would have crowded the cleaner delegation/state thread.
- Rowboat is still a good local-memory artifact, but it is a ledger repeat and did not beat today’s fresher failure and traceability sources.