# Verifiable Loops Win: Codex Testing, GLM 5.2, and Claude Artifacts *By Coding Agents Alpha Tracker β€’ June 22, 2026* The strongest signal today is that coding-agent loops only become useful when the task is verifiable. This brief covers copyable Codex workflows, Riley Brown's GLM 5.2 and Record & Replay demos, Claude Code Artifacts, and a same-day hype check on Sakana Fugu. ## πŸ”₯ TOP SIGNAL Today's clearest pattern: **agent loops only get reliable when the task is verifiable**. Romain Huet says coding is the right proving ground because long tasks can be checked with tests [^1], ThePrimeagen's checklist for successful loops is defined inputs/outputs, clear success/failure, repeatability, and observability [^2], and Armin Ronacher says that without that structure loops still mostly hold up for review/research rather than medium-sized implementation [^3]. Tom Osman and Greg Brockman's Codex workflow is the practical template: generate canonical user stories for every feature, test them, log errors, fix them, then retestβ€”with a human still reviewing PRs before merge [^4][^5][^1]. ## ⚑ TRY THIS - **Run Codex as a full feature-coverage loop (Tom Osman via Greg Brockman).** Point it at an existing app and give it an explicit end state: > `/goal go over every single feature in this app create a user story with expected behaviour based on the code keep a single canonical spreadsheet tracking the features status` > `- when done switch loop to testing every user story and documenting all errors` > `- when done fix every logistical error or ux error` > `- test every user behaviour again post fix` [^4] Greg Brockman highlighted this as Codex for testing every feature, and Tom says it can work through hundreds of user stories automatically [^5][^4]. It also fits ThePrimeagen's loop criteria: defined outcome, clear success/failure, repeatable, observable [^2]. - **Force a second opinion after API design (Theo).** Add this to your Codex first loop: > `When you are done designing the API, get a second opinion from Opus with 'claude -p'` [^6] Theo says this has significantly improved the code quality he gets from OpenAI models [^6]. Good default whenever the agent is making architectural calls. - **Turn a manual browser task into a reusable skill (Riley Brown).** In Codex, use the Record and Replay plugin, say `Please make a skill called [Name]`, perform the workflow on screen, stop recording, and Codex turns it into a slash command you can invoke later like `/manual tweet draft` [^7]. Riley's demo shows recordings up to 30 minutes, which makes this useful for real UI chores, not just toy clicks [^7]. - **Use an agent as a backlog analyst, not just a coder (Geoffrey Huntley).** Give the agent `gh cli` access and ask it to generate a markdown report of the top unresolved issues, with columns for problem description, platform, upvotes, and age, and a linked LLM summary plus proposed resolution for each row [^8]. Huntley's concrete example targets the top 250 unresolved NixOS/nix issues in a file called `nixos-nix.md` [^8]. ## πŸ“‘ WHAT SHIPPED - **GLM 5.2 is now a serious Cursor candidate via OpenRouter.** Riley Brown's setup: in Cursor go to Settings β†’ Models β†’ API keys, enable custom, override the OpenAI base URL with the OpenRouter endpoint, then add `z-ai/glm-5.2` as a custom model [^7]. In Riley's own tests, GLM 5.2 one-shotted a Trello-style app with DB/auth via Convex, built and ran a landing page locally, and handled Notion/Slack agent tasks comparably to Opus 4.8; he also says it feels close to GPT 5.5 / Opus 4.8 overall [^7]. - **Claude Code Artifacts.** Claude Code can now generate shareable interactive mini-apps/artifacts with their own links, giving teams something concrete to review and pass around [^7]. - **Codex stack openness got clearer.** Romain Huet says the Codex CLI, full harness, and server are open source on GitHub; the Codex app can also run open-source models, and he says OpenAI uses Codex across the company, including non-engineers [^1]. - **Temporary deploys for AI-built apps are practical now.** Simon Willison had GPT-5.5 xhigh in Codex Desktop build [cloudflare-redirect-resolver](https://github.com/simonw/cloudflare-redirect-resolver), then deployed it with `npx wrangler deploy --temporary`; Cloudflare kept the ephemeral Workers project live for 60 minutes, and Simon says the temporary deployment worked as advertised [^9]. - **Sakana Fugu launched with an immediate reality check.** Sakana introduced Fugu as a full multi-agent orchestration system behind a single model API and says Fugu Ultra matches Fable and Mythos; it is available at [sakana.ai/fugu](https://sakana.ai/fugu) [^10]. Riley Brown's first design-task test did not finish before daily limits kicked in [^11]. ## 🎬 GO DEEPER - **8:41–12:11 β€” Riley Brown on Record & Replay β†’ Codex skill.** The most copyable walkthrough in today's batch: record a real browser workflow, stop capture, then call it later as a slash command [^7]. [![AI Agents Just Changed Forever: GLM 5.2, Codex Skills, Claude & Cursor](https://img.youtube.com/vi/Nbynj-mKcNI/hqdefault.jpg)](https://youtube.com/watch?v=Nbynj-mKcNI&t=520) *AI Agents Just Changed Forever: GLM 5.2, Codex Skills, Claude & Cursor (8:40)* - **3:34–4:29 β€” Romain Huet on why coding is the first real agent harness.** Short clip, big idea: long-running agents improve fastest where work can be verified by tests and tools [^1]. [![Quel est le futur de l'IA ?](https://img.youtube.com/vi/MxrV8HW2V6w/hqdefault.jpg)](https://youtube.com/watch?v=MxrV8HW2V6w&t=214) *Quel est le futur de l'IA ? (3:34)* - **Repo study β€” Simon Willison's [cloudflare-redirect-resolver](https://github.com/simonw/cloudflare-redirect-resolver) and [build gist](https://gist.github.com/simonw/264bd6b8a39fc34c91c9c867454c64b9).** Small, concrete, and deployed: a good example of using Codex Desktop to build a utility app and ship it to a temporary environment for real validation [^9]. *Editorial take: the durable edge right now is not 'more agents'β€”it is better harnesses: explicit goals, clear success criteria, repeatable environments, verifiable tests, and human review at merge time [^4][^2][^1].* --- ### Sources [^1]: [Quel est le futur de l'IA ?](https://www.youtube.com/watch?v=MxrV8HW2V6w) [^2]: [𝕏 post by @ThePrimeagen](https://x.com/ThePrimeagen/status/2068735111256998074) [^3]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2068657029443363136) [^4]: [𝕏 post by @tomosman](https://x.com/tomosman/status/2068692611334893582) [^5]: [𝕏 post by @gdb](https://x.com/gdb/status/2068761809318990054) [^6]: [𝕏 post by @theo](https://x.com/theo/status/2068595585121484866) [^7]: [AI Agents Just Changed Forever: GLM 5.2, Codex Skills, Claude & Cursor](https://www.youtube.com/watch?v=Nbynj-mKcNI) [^8]: [𝕏 post by @GeoffreyHuntley](https://x.com/GeoffreyHuntley/status/2068744731136499978) [^9]: [Temporary Cloudflare Accounts for AI agents](https://simonwillison.net/2026/Jun/21/temporary-cloudflare-accounts) [^10]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2068861630327443966) [^11]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2068892705497686470)