# Coding-Agent Loops Need Contracts, Not Just Prompts *By Coding Agents Alpha Tracker • July 14, 2026* Today’s strongest practical pattern is the durable coding-agent loop: a clear contract, persistent state, cost-aware triggers, and verification that returns feedback to the original agent. Also included: Codex containment tips, new agent-team and collaboration features, and two implementation walkthroughs worth watching. ## 🔥 TOP SIGNAL **Build agent loops as durable operating systems, not repeated prompts.** Jason Zhou’s team uses a single Markdown document to hold a loop’s goal, autonomy boundaries, SOP, durable state, and append-only run log; after a month automating much of SuperDesign, they report that some loops worked and some did not. [^1][^2] Pair that contract with the right trigger: their cost-conscious “combo” pattern checks a data source with a script first, then wakes the agent only when there is real work. [^1] ## ⚡ TRY THIS - **Put every recurring agent task behind a `LOOP.md`.** Define: (1) a finish line, (2) what the agent may ship or merge alone versus what must escalate, (3) a repeatable SOP, (4) a compact current-state section for hypotheses/backlog/follow-ups, and (5) an append-only log. For scheduled work, run a cheap script first; if it finds no changes, exit without invoking the model. Jason Zhou’s team uses this structure for loops such as daily documentation-drift checks and automated React issue cleanup. [^1] - **Return review findings to the author-agent, then require evidence.** Peter Steinberger’s PR/commit skill invokes an available review CLI, but sends the findings back to the original coding session—the one that understands the task’s constraints—rather than asking a fresh reviewer to blindly patch them. Have that session record accepted or rejected review decisions in the PR description, and define repository invariants in `agent.md` so reviewers do not “fix” intentional behavior. For UI or setup-sensitive changes, give the verifier a fresh-machine box with screenshots and click capability. [^3] - **Contain Codex before it contains your rate limit.** Theo’s current baseline is **High** reasoning, with **Fast mode** and **Ultra** turned off; he characterizes Ultra as an instruction to use more subagents rather than a reasoning level. [^4][^5] Add this guardrail to global `agents.md`: > `Only use subagents if the user explicitly requests them` [^4] Then control autonomy in the task prompt: require a plan and feedback checkpoint, or let the agent build, test, open a PR, handle the first review round, and stop there. Theo’s point: explicit stop conditions are a cleaner control surface than repeatedly changing the harness. [^4] ## 📡 WHAT SHIPPED - **Antigravity Agent Teams:** run `/teamwork-preview` to create specialized subagents that coordinate in the background to plan, build, and verify complex engineering work in parallel. [^6] - **Claude Code Artifacts:** artifacts now support public sharing, multiplayer editing, and creation with Claude Tag. [^7] - **Loopany platform:** SuperDesign open-sourced its internal loop-management tool, which centralizes contracts, state, logs, triggers, and periodic “evolve” runs. Its templates include documentation maintenance, React Doctor, and tech-debt cleanup. [Repo](https://github.com/superdesigndev/loopany-platform) [^2][^1] - **Deep Agents Code on NVIDIA NemoClaw:** LangChain says a single command deploys Deep Agents Code as a governed blueprint using Nemotron 3 Ultra, while retaining control of source, model, and audit trail. [Details](https://www.langchain.com/blog/deep-agents-code-on-nemoclaw-a-governed-blueprint-for-your-most-sensitive-code) [^8] - **Field signal, not a benchmark:** Simon Willison’s `sqlite-utils` 4.0rc2 was “mostly written by Claude Fable” for about $149.25. [^9] ## 🎬 GO DEEPER - **5:05–7:21 — Peter Steinberger on context-aware auto-review.** Watch the implementation detail that matters: run a separate review, but feed the result back to the agent that authored the change so it can judge feedback against the real constraints. [![[한영자막] OpenClaw 창시자 Peter Steinberger가 AI 에이전트를 굴리는 방법](https://img.youtube.com/vi/CRmkxkzrFUA/hqdefault.jpg)](https://youtube.com/watch?v=CRmkxkzrFUA&t=305) *[한영자막] OpenClaw 창시자 Peter Steinberger가 AI 에이전트를 굴리는 방법 (5:05)* - **7:56–9:34 — Jason Zhou on the three-role loop.** A concise production pattern: an orchestrator plans, isolated worktree executors run in parallel, and a verifier attaches test evidence for a human to inspect. [![What I learnt after running loops for 1 month???](https://img.youtube.com/vi/JQ_We_ztxrI/hqdefault.jpg)](https://youtube.com/watch?v=JQ_We_ztxrI&t=476) *What I learnt after running loops for 1 month??? (7:56)* - **Study the Loopany repo.** Its useful contribution is not another agent wrapper—it makes the contract, trigger, state, logs, and iterative improvement cycle first-class artifacts. [Open the repo](https://github.com/superdesigndev/loopany-platform) [^2][^1] *Editorial take: agent autonomy is earned through explicit boundaries, durable state, and verifiable evidence—not by indiscriminately adding more agents.* [^1] --- ### Sources [^1]: [What I learnt after running loops for 1 month???](https://www.youtube.com/watch?v=JQ_We_ztxrI) [^2]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2076626834561925183) [^3]: [\[한영자막\] OpenClaw 창시자 Peter Steinberger가 AI 에이전트를 굴리는 방법](https://www.youtube.com/watch?v=CRmkxkzrFUA) [^4]: [This is absolute chaos...](https://www.youtube.com/watch?v=sKmrLtB47WA) [^5]: [I can't believe they released this](https://www.youtube.com/watch?v=t8hfOyF4ehw) [^6]: [𝕏 post by @antigravity](https://x.com/antigravity/status/2076720528937611363) [^7]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2076789349145092230) [^8]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2076715584092270685) [^9]: [datasette code-frequency chart on GitHub](https://simonwillison.net/2026/Jul/13/datasette-code-frequency)