# Turn Agent Feedback Into Project Memory *By Coding Agents Alpha Tracker • July 26, 2026* Today’s strongest coding-agent pattern is cumulative feedback: preserve corrections as searchable project guidance, then pair parallel execution with bounded scope and independent review. Also covered: signed-in Work browsing, OpenClaw autoreview, and practical Opus 5 routing signals. ## 🔥 TOP SIGNAL **Make agent feedback cumulative.** ThePrimeagen records every agent exchange and commit, then has an agent distill the history into a searchable project `field guide/`; after 10 sessions, he reports that deep feedback fell sharply without changing the model or harness. [^1] Peter Steinberger’s release-prep QA loop applies the same principle at larger scale: bounded parallel work, root-cause fixes, and a continuously updated test report instead of repeated ad-hoc prompting. [^2] ## ⚡ TRY THIS - **Create a retrievable field guide from your corrections.** Start by being exact about the code you want—ThePrimeagen uses voice dictation and nitpicks output. Save every back-and-forth plus the resulting commit hash in Markdown. Periodically ask an agent to review that record and write *durable, generalized* guidance into `field guide/`, with an `init.md` containing links and short descriptions. Let future agents tool-search that directory rather than loading the entire history into context. [^1] - **Run parallel QA with explicit non-negotiables.** Steinberger’s OpenClaw release-prep request: use **12 subagents** split by functionality; run dev gateways on separate ports and reserve some for stress tests; use worktrees and create PRs autonomously; target **200 bugs**; fix root causes rather than patches; allow refactors except across the plugin-SDK boundary; and keep a Markdown test report continuously updated. [^2] This is a replicable pattern: parallelize discovery, but constrain the fix scope and require an auditable artifact. - **Make planning a separate, reviewable phase.** Theo’s workflow is to run Claude Code with Opus 5 at high effort to explore a codebase and propose a plan, have another model independently review that plan, then revise based on the critique. For work intended to merge, he plans to use Opus for implementation and ask Fable and/or 5.6 SOL for a thumbs-up/down review. [^3] - **Feed lint output directly to an agent.** Simon Willison notes that Ruff v0.16.0’s output contains everything a coding agent needs to fix reported problems. Use the linter output as the handoff artifact, then have the agent make and verify the fixes rather than translating errors into a separate prose prompt. [^4] ## 📡 WHAT SHIPPED - **ChatGPT Work agent: signed-in cloud browsing.** You can take over its cloud browser to authenticate on a site, then return control to the agent; the login persists across sessions. Tibo reports the capability is already available in the ChatGPT app. [^5][^6] - **OpenClaw `autoreview` skill:** Steinberger reports a record **66 review rounds** on a complex refactor. The skill definition is public at [openclaw/agent-skills](https://github.com/openclaw/agent-skills/blob/main/skills/autoreview/SKILL.md). [^7] - **Opus 5: strong practitioner results, but watch fit and usage.** Theo calls it the first Anthropic model he has used that follows instructions without repeated constraints, and reports using roughly 20–25% lower cost per task than Fable in his own work. [^3] But reports are not uniform: Armin Ronacher says he likes Opus 5 while finding it expensive and token-hungry, while another early tester reported that existing skills and plugins had to be rebuilt after it argued with instructions and stopped early. [^8][^9] Treat a model upgrade as a harness regression test, not a drop-in swap. - **High-velocity production signal:** Kent C. Dodds reports **135 merged PRs** in kodykoala since Monday, using Cursor with Fable, Grok, and GPT models plus CodeRabbitAI. [^10][^11] ## 🎬 GO DEEPER - **29:20–31:52 — Theo on where Opus 5 fits in a coding stack.** A candid model-routing discussion: Opus’s diligence versus Fable’s taste and SOL’s tendency to overproduce code, plus the rationale for independent review. [^3] [![Opus 5 is my new go-to model](https://img.youtube.com/vi/cIgoqAy_Vs8/hqdefault.jpg)](https://youtube.com/watch?v=cIgoqAy_Vs8&t=1760) *Opus 5 is my new go-to model (29:20)* - **20:16–21:52 — Record a workflow; turn it into a skill.** Riley Brown demonstrates Claude desktop recording screen actions and converting them into a reusable skill—an alternative to writing brittle step-by-step instructions. [^9] [![Opus 5 Is Here… But NEW Claude Voice Is Even Bigger](https://img.youtube.com/vi/Hv8YGUXZY1g/hqdefault.jpg)](https://youtube.com/watch?v=Hv8YGUXZY1g&t=1216) *Opus 5 Is Here… But NEW Claude Voice Is Even Bigger (20:16)* - **Repo to study: OpenClaw’s `autoreview` skill.** The public definition behind a 66-round refactor review is a useful reference for designing persistent critique loops. [Read the skill](https://github.com/openclaw/agent-skills/blob/main/skills/autoreview/SKILL.md). [^7] *Editorial take: the durable advantage is not a single model—it is a workflow that converts failures, reviews, and tool output into compact project memory for the next run.* [^1][^2] --- ### Sources [^1]: [𝕏 post by @ThePrimeagen](https://x.com/ThePrimeagen/status/2081066227619836308) [^2]: [𝕏 post by @steipete](https://x.com/steipete/status/2081169376317932017) [^3]: [Opus 5 is my new go-to model](https://www.youtube.com/watch?v=cIgoqAy_Vs8) [^4]: [Ruff v0.16.0](https://simonwillison.net/2026/Jul/25/ruff) [^5]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2080707685448847418) [^6]: [𝕏 post by @thsottiaux](https://x.com/thsottiaux/status/2081229262452097169) [^7]: [𝕏 post by @steipete](https://x.com/steipete/status/2080899298838098034) [^8]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2081122755018158481) [^9]: [Opus 5 Is Here… But NEW Claude Voice Is Even Bigger](https://www.youtube.com/watch?v=Hv8YGUXZY1g) [^10]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2081002135165231291) [^11]: [𝕏 post by @kentcdodds](https://x.com/kentcdodds/status/2081005013539881005)