# Codex Moves Up to Goal-Level Work While GPT-5.5 Field Reports Stack Up *By Coding Agents Alpha Tracker • May 26, 2026* Today's sharpest shift is from one-shot codegen to long-running, test-backed goal execution. Inside: copyable Codex workflows, skill hygiene, Symfony orchestration, and the clearest GPT-5.5 practitioner signals. ## 🔥 TOP SIGNAL - **The interface is moving from code generation to goal ownership.** Romain Huet says Codex can take a single ambitious `/slashgoal` and run on it for 3-5+ days — Peter Steinberger reportedly let one run for over a week — while the human mostly reviews diffs, comments inline, watches CI, and uses PR review as the control point. Simon Willison's red/green TDD advice is the matching guardrail: keep a test suite that stops agents from breaking old features while they make new changes. [^1][^2] - Huet also says more than 40% of tasks sent to Codex are non-coding automations, a sign that practitioners are already using these tools for broader work than code generation alone. [^1] ## ⚡ TRY THIS - **Run one big goal, then review at the diff/CI layer.** Give Codex a single ambitious objective — e.g. update a neglected legacy codebase and keep going until the goal is reached — and let it work uninterrupted. Then review inline diffs/comments, create the Git commit/push, monitor CI in the app, and turn on Codex Code Review for the PR. Keep a red/green test suite in place so the agent can add new behavior without silently breaking old features. [^1][^2] - **Use an image-first frontend loop.** Ask Codex to generate a design image, iterate on the static result first, then tell GPT-5.5 to implement it. Use the in-app browser to point at elements and give concrete edits like `change this font`, `make it larger`, or `move this to the right`. [^1] - **Trim skill bloat and trust tool-based recall.** Peter Steinberger's instruction when writing skills: tell the agent to `be token efficient` and `relax grammar`, because long skill descriptions get loaded into every context. Then run [`skill-cleaner`](https://github.com/steipete/agent-scripts/blob/main/skills/skill-cleaner/SKILL.md) to find the worst offenders, and let the agent recover forgotten details via tools instead of preloading huge slabs of context — Huet says GPT-5.5 can compact understanding and recall details later when needed. [^3][^1] - **Use read-only audits before cleanup.** KingBootoshi's Codex prompt was `do a FULL read only analysis on my Macbook to help me optimize storage`. Codex reportedly surfaced 500 GB of potential savings and a 116 GB `codex-tui.log` file; the replicable part is the first pass being explicitly read-only. [^4] ## 📡 WHAT SHIPPED - **Symfony** — experimental spec + reference implementation for fleet-of-agents orchestration. Huet frames it as a world where you care about task completion, not staring at code: use Linear as source of truth and monitor which agents are executing or completing each task. [^1] - **Codex stack details** — Huet says the Codex CLI and harness are open source; the harness defines tools, environment, file access, internet access, and MCP servers, and is also part of post-training so the model already knows its tool setup. Adoption signal: he says Peter Steinberger's output doubled after switching to Codex, and he completed 40 open-source projects in one year. [^1] - **skill-cleaner** — Peter Steinberger released a skill to detect overly long skill descriptions. Repo: [skill-cleaner SKILL.md](https://github.com/steipete/agent-scripts/blob/main/skills/skill-cleaner/SKILL.md). [^3] - **GPT-5.5 field signal from DHH.** > "All steering, no handwriting." [^5] DHH says GPT-5.5 has produced more "I can't believe it's this good" coding moments than any model since Opus 4.5, and says it fixed Omarchy 4 Alpha's busted webcam selector in under 2 minutes. [^5][^6] - **Chorus/iMessage build flow** — Riley Brown shared a path to go to [http://chorus.com](http://chorus.com), add Claude to iMessage, and ask it to build an iOS app; he says the method works with any agent. [^7] ## 🎬 GO DEEPER - **11:17-12:51 — Image-first frontend loop.** Best short clip today for anyone building UI with agents: static design first, live browser second, implementation third. [^1] [![AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa](https://img.youtube.com/vi/JPSCTUBfc-M/hqdefault.jpg)](https://youtube.com/watch?v=JPSCTUBfc-M&t=676) *AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa (11:16)* - **13:48-15:00 — Codex as a workbench, not just a code generator.** Inline diff review, Git, CI, PR review, plus the note that 40% of tasks are already non-coding. [^1] [![AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa](https://img.youtube.com/vi/JPSCTUBfc-M/hqdefault.jpg)](https://youtube.com/watch?v=JPSCTUBfc-M&t=827) *AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa (13:47)* - **18:04-19:15 — Why raw context-window talk is fading.** Huet's point: agentic compaction plus tool recall matter more than stuffing everything into the prompt upfront. [^1] [![AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa](https://img.youtube.com/vi/JPSCTUBfc-M/hqdefault.jpg)](https://youtube.com/watch?v=JPSCTUBfc-M&t=1084) *AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa (18:04)* - **24:30-25:09 — Symfony orchestration pattern.** Very short clip, but the idea is durable: Linear as source of truth, multiple agents executing tasks, human supervising outcomes. [^1] [![AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa](https://img.youtube.com/vi/JPSCTUBfc-M/hqdefault.jpg)](https://youtube.com/watch?v=JPSCTUBfc-M&t=1354) *AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa (22:34)* - **Read:** Simon Willison's [red/green TDD for agent code](https://simonwillison.net/guides/agentic-engineering-patterns/red-green-tdd/) — simple rule, high leverage. [^2] - **Read:** Peter Steinberger's [skill-cleaner SKILL.md](https://github.com/steipete/agent-scripts/blob/main/skills/skill-cleaner/SKILL.md) — good example of using a skill to police other skills. [^3] *Editorial take: the durable edge today is ambitious async goals plus tight guardrails — tests, lean skills, and review/CI checkpoints — not longer prompts.* [^1][^2][^3][^1] --- ### Sources [^1]: [AGI, le futur de l'ingénieur, Codex, - OpenAI head of Developer Experience x Hexa](https://www.youtube.com/watch?v=JPSCTUBfc-M) [^2]: [𝕏 post by @simonw](https://x.com/simonw/status/2058992972734341445) [^3]: [𝕏 post by @steipete](https://x.com/steipete/status/2058917897590673525) [^4]: [𝕏 post by @KingBootoshi](https://x.com/KingBootoshi/status/2058943762207006959) [^5]: [𝕏 post by @dhh](https://x.com/dhh/status/2058953269360156783) [^6]: [𝕏 post by @dhh](https://x.com/dhh/status/2058976267626492251) [^7]: [𝕏 post by @rileybrown](https://x.com/rileybrown/status/2059052352875946360)