# Proactive Agent Loops and Auto-Review Enter the Workflow *By Coding Agents Alpha Tracker โ€ข June 12, 2026* Today's useful signal is the operational layer around coding agents: proactive debug loops, external judge agents, auto-review defaults, and transcript-level evals. This brief pulls together the most actionable workflows, tool updates, and clips worth studying. ## ๐Ÿ”ฅ TOP SIGNAL - **Proactive loops are the real step change.** Simon Willison's firsthand Fable transcript is the clearest proof: from a screenshot and a one-line goal, the agent stood up a local server, scratch HTML, browser automation, DOM probes, and screenshot capture to isolate a two-line CSS fixโ€”and Willison says that same relentless proactivity is exactly why unsandboxed agents are a security risk if prompt-injected [^1]. Cursor's multi-agent practitioners described the same broader pattern at scale: plan-first swarms, reminder loops, and separate judge agents that keep long-running tasks from stopping early [^2]. ## โšก TRY THIS - **Steal Simon Willison's browser-debugging escalation ladder.** Start with a screenshot and a one-line objective. If the bug does not reproduce in Playwright across Chrome/Firefox/WebKit, have the agent: run the local dev server with fake env vars; write a scratch HTML page that isolates the behavior; open real Safari/Firefox; use `uv run --with pyobjc-framework-Quartz` plus `screencapture -x -o -l` to capture real browser windows; inject JS to trigger hidden UI flows; and post DOM measurements back to `127.0.0.1:9999` via a tiny CORS `http.server` so the agent can read the results from disk [^1]. > `Write a report in /tmp/automation-report.md where you note down all of the tricks you have used in this session to test against real browsers on my computer, include runnable code examples` [^1] - **Run a plan-first swarm, not a prompt pile.** In Cursor's multi-agent discussion, the advice was to spend most of the session in plan mode until you have a detailed markdown spec, then let a main agent delegate to named sub-agents via messaging scripts, with reminder loops that keep overnight runs anchored on explicit verification steps. For long runs, pass data across compaction cycles by reference instead of summaries [^2]. If you push toward bigger fleets, don't assume GitHub is your control plane: swyx's summary from ai.engineer says runtime and triggers are mostly solved, but coordination is not, and GitHub gets too noisy for hundreds of parallel PRsโ€”use a CLI or messaging gateway instead [^3]. - **Split build vs. review across model families.** A concrete pattern from Cursor's panel: use Claude/Opus 4.7 for implementation or design, keep a GPT-5.5 worker around for review and functional verification, and run several rounds of "thermonuclear review" so different models catch different failure modes [^2]. Simon Willison used the same basic routing in a fresh side project: Claude Fable 5 for the plan, GPT-5.5 xhigh in Codex Desktop for the implementation of a custom Datasette extras explorer [^4]. - **Read trajectories like code review, not scoreboard output.** swyx's Frontier Code discussion argues for hard, real-world tasks plus maintainer-labeled trajectories and rubrics for mergeability, not just passing tests; the concrete checks are simple: did the agent touch the right files, write reasonable tests, and respect style/lint guidelines [^5][^6]. Sourcegraph's 1,281-run study points the same way: failures in large repos show up as repeatable patterns tied to different infra fixes, not random bad luck [^7]. ## ๐Ÿ“ก WHAT SHIPPED - **Cursor Auto-review** โ€” now default for new users; a classifier subagent evaluates actions in context and decides whether to allow, block, or ask for approval. Cursor says evals show **97% accuracy**, with misses mostly near ambiguous edges. Details: [cursor.com/blog/agent-autonomy-auto-review](http://cursor.com/blog/agent-autonomy-auto-review) [^8][^9]. - **Google's Anti Gravity harness, per Logan Kilpatrick** โ€” framed as a shared IDE/web/CLI/SDK + Gemini API managed-agent layer, with the same harness powering Search, Gemini app, Cloud, and AI Studio; use-case variants appear to share roughly **80%** of the base harness [^10]. Usage signals from Logan: Google teams used it to ship mobile and macOS apps faster internally, AI Studio saw roughly **350k Android apps** built since last week, and about **20%** of early apps were games [^10]. - **Gemini coding-model note** โ€” Logan says **Gemini 3.5 Flash** is better at coding than any prior Pro model Google released, and credits the jump to post-training gains rather than a bigger base model [^10]. - **Sourcegraph's large-repo postmortem** โ€” **1,281** agent runs across **40+** open-source repos surfaced five repeatable failure patterns, each pointing to a different infra fix. Read: [Why coding agents fail in large codebases (and what to do about it)](https://sourcegraph.com/blog/why-coding-agents-fail-large-codebases?&utm_medium=social&utm_source=twitter&utm_campaign=agent-effectiveness) [^7]. - **Open-source maintainer loop from Peter Steinberger** โ€” his Codex setup wakes every five minutes, steers work to threads, and combines an orchestrator with triage/autoreview/computer-use skills so some changes can land autonomously. Study the skill files: [maintainer-orchestrator](https://github.com/steipete/agent-scripts/blob/main/skills/maintainer-orchestrator/SKILL.md) and [github-project-triage](https://github.com/steipete/agent-scripts/blob/main/skills/github-project-triage/SKILL.md) [^11]. - **Fresh small-project proof of routing** โ€” Simon Willison shipped a [custom extras API explorer](https://tools.simonwillison.net/datasette-extras-explorer) with Fable 5 handling the plan and GPT-5.5 xhigh handling the implementation; separately, he says Fable spotted and fixed bugs in `asyncinject 0.7` [^4][^12]. ## ๐ŸŽฌ GO DEEPER - **Cursor, 5:19-5:54 โ€” separate judge agents.** Best short explanation of why a worker agent needs an external checker to stop false-done states before they spread [^2]. [![Running 128 Coding Agents at Once](https://img.youtube.com/vi/-jnwTZ789V0/hqdefault.jpg)](https://youtube.com/watch?v=-jnwTZ789V0&t=319) *Running 128 Coding Agents at Once (5:19)* - **Google/Logan, 38:39-40:30 โ€” the model eats the harness.** Useful mental model for why today's bespoke agent scaffolding keeps getting upstreamed into model behavior, and where the next edge moves [^10]. [![Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness](https://img.youtube.com/vi/cMAs8z2dehs/hqdefault.jpg)](https://youtube.com/watch?v=cMAs8z2dehs&t=2318) *Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness (38:38)* - **Frontier Code, 1:01:05-1:05:55 โ€” trajectory reading party.** Best clip today on how to judge agent work beyond pass/fail: wrong-file touches, reward hacking, weak tests, and mergeability [^6]. [![Fable 5 controversy, new Siri with computer use, Gemini Live Translate under 500ms | Jun 11](https://img.youtube.com/vi/Ci4TYEze9Gg/hqdefault.jpg)](https://youtube.com/watch?v=Ci4TYEze9Gg&t=3665) *Fable 5 controversy, new Siri with computer use, Gemini Live Translate under 500ms | Jun 11 (61:05)* - **Study Simon Willison's [full terminal transcript](https://gisthost.github.io/?cc14774f6d37eb67bf089f3ac3925f8f).** It's one of the best public artifacts right now for seeing an agent improvise across Playwright, Safari, DOM instrumentation, screenshots, and local servers in a real debugging task [^1]. - **Study steipete's skill files.** If you want a concrete maintainer loop instead of general agent talk, the orchestrator and triage skills show how he structures autonomous repo maintenance around threads and review gates [^11]. [maintainer-orchestrator](https://github.com/steipete/agent-scripts/blob/main/skills/maintainer-orchestrator/SKILL.md) ยท [github-project-triage](https://github.com/steipete/agent-scripts/blob/main/skills/github-project-triage/SKILL.md) *Editorial take: the edge is moving from raw codegen to agent control systemsโ€”loops, judges, artifact capture, and transcript review are what make longer autonomy usable.* --- ### Sources [^1]: [Claude Fable is relentlessly proactive](https://simonwillison.net/2026/Jun/11/fable-is-relentlessly-proactive) [^2]: [Running 128 Coding Agents at Once](https://www.youtube.com/watch?v=-jnwTZ789V0) [^3]: [๐• post by @aiDotEngineer](https://x.com/aiDotEngineer/status/2065175479343284421) [^4]: [datasette 1.0a33](https://simonwillison.net/2026/Jun/11/datasette) [^5]: [LIVE - Fable/Mythos 5 is HERE , SIRI AI finally?, Frontier code w/ Swyx & more AI news](https://www.youtube.com/watch?v=YVUm0KUMNLc) [^6]: [Fable 5 controversy, new Siri with computer use, Gemini Live Translate under 500ms | Jun 11](https://www.youtube.com/watch?v=Ci4TYEze9Gg) [^7]: [๐• post by @Sourcegraph](https://x.com/Sourcegraph/status/2065212260637868389) [^8]: [๐• post by @cursor_ai](https://x.com/cursor_ai/status/2065137803084857845) [^9]: [๐• post by @cursor_ai](https://x.com/cursor_ai/status/2065137805685334463) [^10]: [Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness](https://www.youtube.com/watch?v=cMAs8z2dehs) [^11]: [๐• post by @steipete](https://x.com/steipete/status/2064998499780084154) [^12]: [asyncinject 0.7](https://simonwillison.net/2026/Jun/11/asyncinject)