Best tweets about OpenAI Codex
Explore the best tweets about OpenAI Codex, including coding-agent workflows, CLI use, delegated tasks, reviews, and engineering results. Updated weekly.
Real Codex tasks and software-engineering outcomes, excluding unrelated historical or publishing references to codex.
Best OpenAI Codex tweets
Ranked 01–50
@OpenAI ·
Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app.
Watch video
@om_patel5 ·
THIS GUY MADE CLAUDE CODE AND CODEX TALK TO EACH OTHER not by switching between them, they literally talk to each other about your code in real time you give it a task. it spawns a Claude agent and a Codex agent in separate sessions. they share findings, challenge each other, and produce a combined result. he pointed them at an Express API with intentional security issues. within minutes they found SQL injection, hardcoded secrets, missing auth, and weak JWT config. one catches something and the other confirms it and adds context. they divide the work on their own for faster speed one command: /collab "review auth.js for security issues" that's it. you watch the conversation happen in real time. AND its open source with new models being integrated crazy to think that this was entirely vibe coded
Watch video
@mark_k ·
Greg Brockman explains the plans for the @OpenAI SuperApp, fusing ChatGPT and Codex into a single UI. The Codex app is going to be the base for this mega application, deprecating the ChatGPT app.
Watch video
@arrakis_ai ·
🚨 BREAKING: Codex in Chrome just leaked!!! Okay, the leak was a joke. But the tool is real. I used the open-source Codex App Server and ChatGPT OAuth to build a fully functional Codex Extension directly inside Google Chrome. Here is what it can do right out of the box: → Image Editing: Edit the exact images you are currently viewing on the web. → Contextual AI: Ask questions based on your active tabs and actual search history. → Page Analysis: Extract and process information directly from your current webpage. → Browser Automation: Full browser control and execution. No more copying and pasting URLs or screenshots into a separate window. The AI lives in your browser, sees what you see, and acts on it. The best part? It is 100% open-source. You can grab the code, install the extension, and run it right now for free.
Watch video
@petergyang ·
"We write very few specs on the Codex team. We're talking 10 bullet points and that's it." Here's my new episode with @embirico and @romainhuet where they gave me an inside look at how OpenAI's Codex team operates: → Live demo: Building in seconds with Spark → How the team built the beautiful Codex app → How they ship without traditional specs and roadmaps Some quotes from Alex and Romain: "The fewer people you need in a room to do anything, the more pure every decision is." "Our designers write more code now than was written by an engineer six months ago." "I'm much less likely to read someone's resume than their ideas and what they've built." Thanks to our sponsors: @meetgranola: The best AI meeting notes app I've ever used https://t.co/MNToIh5WTm @linear: The AI agent platform for modern teams https://t.co/lI40xrrDsr 📌 Watch now: https://t.co/2lbvdTiQUl
Watch video
@xanderai ·
In the few weeks since Claude Code's agent teams and Codex's multi agents features came out, I've personally spent ~50B tokens between the two of them. Some observations: 1) codex cli UI is much, much smoother. The claude code cli at some length of conversation history / number of running background tasks becomes incredibly slow and sagging. Difficult even to type anything. 2) codex has no mechanism by which the main/leader agent wakes up to respond to sub agents' results. Claude code has a communication mechanism where sub agents send messages to the leader when they have interesting results. These messages wake up the leader and it thinks and takes actions to make sure the whole endeavor is on track. codex's main agent just sleeps forever after initial setup, requiring me to prod it regularly. 3) claude code's tmux integration is a limitation. I've seen it unable to spawn new agents simply because tmux refuses to create more panes: "no space for new pane" - required to use in-process agents, and no way to change this after starting. 4) Claude code has been much harder to get rate limit increases on. I have ~100 simultaneous agents that regularly all get shut down due to hitting overloaded or per minute rate limit errors. I've encountered hundreds if not thousands of rate limit errors that are never reported in the Anthropic Console. 5) Claude code *really* wants to end sessions after a certain amount of time. There's no amount of prompting I can do to prevent it from doing this. There's either some training or prompt injection that makes it want to shut down prematurely. Telling it to stop that and keep going works for ~20min only. Claude loves to produce a "FINAL report" while there are still hundreds of productive tasks in process. 6) codex's agent names are non-descriptive and useless. Hegel, Lovelace, Feynman, Pauli, Curie, ... Come on. I greatly prefer claude code's descriptive agent names: advisor, perf, node-1, node-2, researcher, X-debug, etc. Less cute. More utility. I have no clue which agent I need to talk to when looking at the agent list in codex. 7) In claude code it is difficult and even non-deterministic to interrupt the main agent's currently blocking task. The main agent should *never* sleep on its blocking thread. And yet all the time I have to send messages and hit escape back and forth until it gets interrupted. In codex this is a 100% reliable single Esc key press. #1 is the biggest reason I prefer codex for this. #2 is the biggest reason I prefer claude code for this. Will make immediate use of any fixes for these issues @thsottiaux @bcherny
@mark_k ·
It's CODEX THURSDAY and OpenAI came through! 🔥 Codex app 26.616 changes: • Added Record & Replay on macOS, which turns a demonstrated workflow into a reusable skill. • Record & Replay is not available in the EU at launch. • Record & Replay requires Computer Use to be enabled by the user or admin. • Added bulk actions to automation run history, so runs can be marked as read or archived in bulk. • Added new deep links for managing SSH connections. • Improved Browser Use so visible-tab routing and annotations persist when a draft browser session moves to the server. • Additional performance improvements and bug fixes.
@marvinvista ·
Codex as Chief of Staff For knowledge work: Plugins enable Codex to do most of my knowledge work. It reads and responds to emails and Slack. It builds models in Sheets and handles all of my note taking for me. Codex is quickly becoming my assistant. It understands me and what I’m doing enough to help get real work done. - @trekedge, https://t.co/NJiXNtQmjX For dinner planning: Our incredible comms leader, @lindsmccallum, planned a closed door dinner for the Codex team in a fraction of the time it would normally take - thanks to Codex.She used the Codex App to: - compile the invite list - send out invitations - hourly scan of her emails to update RSVP status - populate a doc with bios on every attendee - create a mini app to plan the seating chart - @rohanvarma, https://t.co/p6LpVKNBEX Inbox triage: Getting codex's help to me climb from under my inbox mountain - @boazbaraktcs, https://t.co/JXz57p8nRq Automation density: I've been using plug-ins a ton internally. I have about 58 automations and 30 plug-ins and I've automated everything except the part where I have to come up with ideas and talk to people - @jxnlco, https://t.co/w0K2GaKzzl
@cerebras ·
what can you build with gpt-5.3-codex-spark? @jxnlco from @OpenAI demos 3 real workflows — ones you can set up yourself inside the Codex app to help you spend less time on overhead and more time building. 00:09 – what is gpt-5.3-codex-spark? 00:25 – workflow 1: multi-agent daily briefing from slack, drive & meets 01:06 – workflow 2: automated PR review 01:31 – workflow 3: real-time interactive coding 02:56 – what speed changes, and what's coming next
Watch video
@bradmillscan ·
Did you know with OpenClaw you can talk to ChatGPT Codex on Telegram kinda like Claude's Dispatch? It inherits workspace visibility and it runs outside of the openclaw subagent sandbox. So it's pretty powerful - but it's not an isolated agent. 1 Setup Telegram Topics 2 Enable Telegram ACP thread spawning 3 set channels.telegram.threadBindings.spawnAcpSessions = true 4 add an ACP agent entry for Codex 5 pin the agent to a topic with channels.telegram.groups.<chat>.topics.<id>.agentId (If you're curious, it does have a session but it does not have auto-compaction or write to the lossless-claw DAG because that is for the openclaw pi runs and codex is CLI backend.) I'm using mine for tech support. Experimenting with it, might not keep it around.
@AbdelStark ·
I am stocked to announce that I won the @OpenAIDevs Codex x Mollie Hacka Worldwide Hackathon in Paris. 60+ builders, every one of us working solo, one day to ship. I built mine around a single question: who gets to own intelligence? The default answer is scary. You hand your data to a handful of labs, they train the model, they own it, and you rent back a thin slice of what your own data made possible. That is the bargain on the table today. I do not accept it. So I built Lensemble: a Tapestry like distributed training platform for JEPA based World Models. What does it enable: World Models that a community improves together, keeps sovereign, and co-owns. Two bets sit underneath it. First, the paradigm. Language models predict the next token. Powerful for text, a dead end for the physical world. A robot does not need to autocomplete sentences, it needs to predict what happens next in the world. That is what JEPA does: it learns by predicting representations instead of pixels or tokens. I am convinced world models are the most underrated paradigm in AI right now, and the closest thing we have to a ChatGPT moment for robotics. Second, the politics. Your raw trajectories never leave your machine. Each participant trains locally against a shared protocol and ships only an update, never the data. A federated round folds those updates into one shared world model, a LeWorldModel based model, and the gain is measured, not claimed: a 12k-parameter adapter on a frozen backbone, held-out prediction error down about 12 percent, the model measurably less surprised by the world. Then the upside is split by contribution weight, so the people who improved the model own a share of what it earns. This is the thesis behind Project Tapestry, the AI Alliance and Yann LeCun's push for federated, sovereign frontier AI, carried into world models and robotics. Call it Tapestry for the physical world. All of it built solo, in a single day, with Codex as my pair the whole way. Thank you to OpenAI Codex and Mollie for backing builders who ship real things, and to @borvibe and the organizing crew for the room and the standard you set. Intelligence the world improves, and the world owns. That is the future I want for my kids, and the one I will keep building.
Watch video
@VaibhavSisinty ·
GPT 5.5 Changes Everything 🤯 I asked OpenAI Codex to build a full Mac app while I slept. No code. No dev team. No late-night debugging. One prompt. One goal. An AI coding agent that kept moving until the job was done. GPT-5.5 is here. Codex is here. Claude Opus 4.7 is here. AI agents are here. The question is no longer who writes the better prompt. The question is who lets AI execute first.
Watch video
@emanueledpt ·
After creating Remodex: Codex remote control for iOS I decided to chill a bit by building a browser sidebar inside @theo 's T3 Code You can open it with Cmd+Shift+B This software is so easy to "manipulate" and build upon Congrats to all the guys that put the work in it because it is great I certainly will be building more things with it For this one I decided to switch from the Codex App to the Codex CLI too just to see how it works And boy it was so good
Watch video
@TheRohanVarma ·
Our incredible comms leader, @lindsmccallum , planned a closed door dinner for the Codex team in a fraction of the time it would normally take - thanks to Codex. She used the Codex App to: - compile the invite list - send out invitations - hourly scan of her emails to update RSVP status - populate a doc with bios on every attendee - create a mini app to plan the seating chart Things Codex didn’t do: - make the sushi we ate (soon) Lindsay has never coded before. With Codex, she is a builder. Codex is the interface for personalized software.
@free_ai_guides ·
7 mistakes I see people make with OpenAI Codex: 1. No AGENTS. md file. Codex guesses your stack every session. 2. One thread per project. Context rots. Codex starts repeating itself. 3. Using GPT-5.4 for everything. Simple tasks burn quota for no reason. 4. No skills installed. You re-explain the same patterns every time. 5. Watching it step by step. It's built for parallel work, not babysitting. 6. Zero MCP servers connected. You're copy-pasting what Codex could pull itself. 7. Skipping verification. Confident output doesn't mean correct output. Fixes for all 7 in the infographic below.
@Hesamation ·
OpenAI reports: Codex active users grew 5x in the first half of 2026. this is from OpenAI's paper released today, based on the data from Codex. main takeaways: > non-dev Codex usage is exploding. Since Aug 2025, it’s up ~189x in orgs and ~137x for individuals > inside OpenAI, Codex is now 99.8% of their output tokens and ChatGPT only 0.2% > individual users are still mostly in ChatGPT: Codex is only 16.5% of their output tokens > organizations are using Codex like no one else. Codex went from almost nothing in Aug 2025 to 63.3% of output tokens now > long horizon tasks are taking over. 1h+ human-task prompts went from 35.4% of users in Dec to 70.2% in May > 8h+ human-task prompts went from 2.1% in Dec to 25.6% in May >skills are being adopted fast. usage went from 5.4% in Mar to 26.6% in Jun > OpenAI’s 99th-percentile users are running ~71 hours of agent work per day, they’re managing multiple agents in parallel
@aiwithjainam ·
Someone just gave OpenAI Codex its own App Store. 136+ subagents. Each one a specialist. Each one isolated. It's called Awesome Codex Subagents. → Code review agent with its own context window → Security auditor hunting only vulnerabilities → Debugger tracing root causes from scratch → Docs writer that never pollutes your main thread Drop .toml files into ~/.codex/agents/ and they're live globally. No more one AI juggling 50 things and forgetting half. 100% Opensource. Link in comments.
@alexabelonix ·
OpenAI Codex app is now on Windows. You get the full Codex app experience on Windows with a native agent sandbox and support for Windows developer environments in PowerShell.
Watch video
@shao__meng ·
OpenAI Codex - Agent Skills Best Practices 只有四句话： 1. Keep each skill focused on one job. 2. Prefer instructions over scripts unless you need deterministic behavior or external tooling. 3. Write imperative steps with explicit inputs and outputs. 4. Test prompts against the skill description to confirm the right trigger behavior. https://t.co/KNb0CCsJxc 1. Keep each skill focused on one job Skill 职责边界要单一： · description 越单一，模型越容易判断“这次该不该用这个 skill” · 一个 skill 做太多事，会出现触发不准 · 维护时也更容易改坏别的场景 举个例子： · 不推荐：fullstack-helper，既管数据库，又管前端，又管部署 · 推荐：postgres-query-tuning、playwright-ui-debug、release-note-writer 判断标准： · 你能不能用一句话说清“它只解决什么问题” · 用户提一个任务时，你能不能很快判断“该用它”还是“不该用它” 2. Prefer instructions over scripts unless you need deterministic behavior or external tooling 默认先写清楚步骤，不要一上来就写脚本，为什么： · 说明文字更灵活，模型可以根据现场情况调整 · 脚本需要维护、兼容环境、处理错误，成本更高 · 很多任务本来只是“怎么做”的问题，不一定需要固定程序 文档也给了两个例外： · 你需要确定性行为 例如输出格式必须严格一致、每次都要按同样流程执行 · 你需要外部工具 例如调用 ffmpeg、数据库客户端、测试命令、浏览器自动化工具 可以这样理解： · 能靠明确步骤完成，就先用文字 · 只有在“必须固定”或“必须调用工具”时，再加脚本 3. Write imperative steps with explicit inputs and outputs 用命令式写法，输入输出写明白： · 用“先做 A，再做 B，然后检查 C” · 不要写成背景介绍、经验分享、长篇解释 输入输出写明白指的是： · 输入是什么：用户给什么、文件在哪、参数是什么 · 输出是什么：生成什么文件、返回什么格式、最终要交付什么结果 不好的写法： “分析代码库并给出建议。” 更好的写法： “读取 package.json 和 src/ 下的构建配置，找出构建失败原因。输出一个按严重程度排序的问题列表，并给出最小修改方案。” 4. Test prompts against the skill description to confirm the right trigger behavior. Codex 会根据 description 做隐式匹配，所以你写完 skill 后，不能只看内容是否正确，还要测“会不会在对的时候被选中，在不对的时候不被选中”。 应该怎么测： · 用 3 到 5 个“应该触发”的提示词试一下 · 再用 3 到 5 个“看起来有点像，但其实不该触发”的提示词试一下 真正要验证的是： · description 是否太宽，导致误触发 · description 是否太窄，导致该触发时没触发 · skill 名字、说明、步骤之间是否一致 最后 - 给咱们的参考：一个好的 skill，应该同时满足 · 触发边界清楚 · 步骤可执行 · 结果可检查 · 只有在确实需要时才引入脚本
@rstormsf ·
Vibe coded a small app in the new Codex app by @OpenAI. 🪄 Spec: A standalone HTML tool for encrypting/decrypting seed phrases with multiple passphrases. To be honest, I’m curious how Codex actually makes its security decisions under the hood - is a "vibe coded" security tool actually useful, or is it a liability? Checking the source now. 🔐🤔 Source: https://t.co/Xv1HzMqkKj Demo: https://t.co/acINV9IvCh
@rohitdotmittal ·
I just asked Codex to watch a Loom video about a product improvement, find the relevant parts, take screenshots, and create updates about the product. Codex on 5.6 Sol used the Codex app, the browser, and the MCP servers across 3 different tools/services to complete the task. And it did it way better than Claude Fable 5. For browser automation, Codex 5.6 Sol completely beat Claude Fable 5. More bullish on using AI for future workflow enhancements.
@DeryaTR_ ·
With OpenAI Codex automations, I built HTML file that update daily & list all the emails in my inboxes, organized by priority & groupings. Each email also includes direct links to open it, along with archive & delete buttons, so I can manage messages directly from the HTML file!
@RoundtableSpace ·
Someone figured out how to run Claude Fable 5 and GPT-5.6 together inside ChatGPT Codex and the setup takes 5 minutes. Two commands get you started: → codex plugin marketplace add Cjbuilds/Codex-Orchestration → codex plugin add codex-orchestration@codex-orchestration From there you set an advisor model for the expensive thinking and an executor model for the cheaper implementation. Paste the workflow prompt, describe the task and both models handle their side automatically. Fable 5 plans. A cheaper model executes. You get the best of both without paying flagship prices for every token.
Watch video
@Suryanshti777 ·
Holy shit...OpenAI just killed its most successful app. And it's the smartest thing they've done all year. The Codex app — 5 million weekly users, the fastest-growing dev tool in AI — is gone. As of yesterday, it IS the ChatGPT desktop app. Why would you kill something that's working? Because of one stat buried in their launch post: 1 MILLION people were using Codex for work that had nothing to do with code. Finance teams. Sales teams. Marketers. All quietly using a "coding agent" because it was the only AI that actually finished tasks instead of chatting about them. OpenAI saw it and made the call: the agent isn't a dev tool. It's the product. So now Chat, Work, and Codex live in one desktop app — on every plan, including free — and it can use your files, your apps, and your browser. The chatbot era ended yesterday. Nobody sent a press release saying so, because the press release was about benchmarks. But watch what people build this month.
@phuctm97 ·
ChatGPT/Codex Computer Use is the most underrated piece of software in the AI era. It's insanely good and unblocks essentially infinite amount of things that AI can do for me: • Test my web/desktop/mobile app end-to-end • Test any integrations with other apps • Use any other apps that don't have an API/CLI
@feifei_qiu ·
I'm liking ChatGPT/Codex more and more. Everything is in one place: chat, coding, GitHub, and connectors. A small example: yesterday I wanted to search my Outlook emails. I first tried connecting Claude Code, but there wasn't a direct connector. I tried custom one, but it only supported organizational Outlook accounts, not personal ones. Then I tried Codex. One click away, and I was searching my emails. User experience matters but got ignored or deprioritized by lots of companies Plus, ChatGPT leading image and coding model, Right now, I think ChatGPT is winning
@frog_omo ·
openai codex is the most misunderstood tool in AI right now. most people think it's just "chatgpt for code." it's not. it's an autonomous coding agent that reads your entire repo, writes code across multiple files, runs tests, iterates on failures, and opens pull requests — while you do something else. here's everything you need to know: what codex actually is: you describe a task in plain english → codex plans, writes code, installs dependencies, runs your test suite, iterates on failures, and returns a reviewable PR. it runs in isolated cloud sandboxes. no internet access by default (prevents supply-chain attacks). tasks take 1-30 minutes. this isn't autocomplete. it's delegation. how to use it: → CLI: npm i -g @openai/codex, then run "codex" in your terminal → cloud: https://t.co/6LT2EArDPR — connect github, configure your repo, launch tasks → IDE: vs code, cursor, jetbrains extensions → desktop app: macos/windows command center with parallel threads → github bot: tag @codex on any PR or issue the secret weapon: AGENTS.md this is a markdown file in your repo that tells codex how your project works — setup commands, test commands, conventions, architecture rules. codex reads it before every task. this is the single biggest quality lever. write it before your first task. commands first, prose second. best use cases (where codex shines): → bug fixes from issue descriptions → refactoring, renaming, code migrations → test coverage expansion → dependency upgrades → documentation from code diffs → automated PR code review (openai's most-used internal feature) → scaffolding new features → maintenance backlog you never get to openai built the sora android app in 28 days using codex. in one experiment, codex ran 25 hours straight, used 13M tokens, and generated 30k lines of code. what it's NOT good for: → interactive/exploratory debugging (use cursor or claude code) → frontend/visual work where you need to see updates live → high-stakes first-pass correctness (claude opus 4.7 is safer here) → loose prompts — it will install packages and edit files you didn't mention the "act, don't ask" default creates sprawling diffs when assumptions are wrong. pricing: bundled into chatgpt plans — no standalone subscription. → plus ($20/mo): solid daily usage for individuals → pro ($200/mo): heavy usage, ~20x plus limits → business: $20/seat annually, or codex-only pay-as-you-go pro tip: plus at $20/mo is more token-efficient than claude pro at the same price for batch work. the expert consensus: use both codex and claude code. → codex: cheap, autonomous, parallel batch work — PR review, maintenance backlog, test coverage → claude code: interactive, high-stakes, first-pass-correct edits and architecture "claude generates, codex reviews" is a popular pattern. or vice versa. how to start (this week): 1. if you have chatgpt plus, codex is free at the margin, try it 2. install the CLI, authenticate with your chatgpt account 3. run a read-only task first: "explain this project's structure" 4. write AGENTS.md (setup, test, lint commands) 5. batch your maintenance backlog into clear issues 6. delegate in parallel, review and merge never merge without running your own tests. codex can be confidently wrong. the 5M+ weekly users aren't wrong. this is the real shift.
@RoundtableSpace ·
THE CODEX APP SERVER LETS YOU INJECT CODEX-LEVEL INTELLIGENCE INTO ANY PLATFORM USING YOUR CHATGPT ACCOUNT. SOMEONE EMBEDDED IT DIRECTLY INTO CHROME AND IT WORKS FLAWLESSLY. 100% OPEN SOURCE.
Watch video
@burkov ·
A paradigm shift: OpenAI Codex is now part of ChapterPal's business logic. A significant part of what I do when I work on ChapterPal is to run a conversion from LaTeX into Markdown and see what failed. LaTeX is very old and versatile, so some scientists use very crazy tweaks in their LaTeX sources to achieve a desired visual effect in the output PDF which fail at conversion in ChapterPal. Previously, I was in the loop: 1. Me: run a conversion 2. Me: validate visually 3. Me: instruct a coding agent to fix the issues 4. Coding agent: makes fixes, run tests 5. Me: Go to 1. Now me is out of the loop and all happens within a cloud run. I can only observe the process in the admin's UI. So, the conversion code self-writes while I sleep and the conversion almost never fails because a failed conversion is analyzed by Codex and the conversion code is fixed.
@JulianGoldieSEO ·
OLLAMA + CODEX APP JUST BROKE LOCAL AI CODING You can now run OpenAI’s desktop coding agent on models sitting on your own laptop. No subscription. No cloud code sharing. No waiting in line. What Changed: → Ollama 0.24 adds official support for the Codex app → Codex can now run on local models like Qwen 3, Gemma 4, GLM, DeepSeek, and Kimi → You launch it with one command and pick your model from the dropdown Why This Is Huge: ✓ Codex gets parallel work trees for isolated coding tasks ✓ Built-in browser lets you point at UI elements and ask for edits ✓ Review mode lets you diff, comment, branch, commit, and merge in one place Best Use Cases: → Private code refactoring offline → Coding on planes or bad Wi-Fi → Building small internal tools in hours → Automated code review for solo builders → Learning to code without burning usage limits Model Picks: ✓ 16GB RAM: Qwen 3 14B ✓ Lower RAM: Qwen 3 8B or Gemma 4 4B ✓ Bigger jobs: Kimi K2.6 through Ollama Cloud The lesson: Local AI coding is no longer just a nerdy terminal trick. It is turning into a real agent workflow you can actually build with.
Watch video
@charliejhills ·
Claude Code or Codex is the wrong question. You can (actually) run both. Ask these 4 questions instead: Both read your repo, edit files and run commands. The defaults are what actually separate them. 1. Where do you want to work? Claude Code centres on the terminal. Codex spans app, IDE, cloud and GitHub by default. Pick the surface you already work in. 2. How do you split up the work? Both can run agents in parallel. Claude Code delegates to subagents in one session. Codex fans work out to threads you supervise. One nests the work. One lines it up for you. 3. How do you encode your workflow? Claude Code defaults to CLAUDE.md. Codex defaults to AGENTS.md. Both read MCP. Write the rules your team will maintain. 4. Where should long-running work happen? Claude Code runs headless in your shell and CI. Codex leans on cloud jobs and a review queue. Codex runs headless too. I use codex exec. Send the long job to CI or the cloud, not your laptop. Then pick: ✦ If you live in the terminal, run Claude Code. ✦ If you want more surfaces, run Codex. ✦ If you want pairing plus background runs, run both. I run both. Claude Code builds, Codex attacks it. OpenAI shipped Codex as a Claude Code plugin. The integration only runs one way. Install it inside Claude Code: /plugin marketplace add openai/codex-plugin-cc /plugin install codex@openai-codex Then run /codex:adversarial-review Which model, not which agent? Claude 5 vs ChatGPT 5.6 https://t.co/ZeHn9pSiFc Repost ♻️ to help someone stop choosing. P.S. Which coding agent do you use?
Watch video
@rohanpaul_ai ·
OpenAI published engineering deep dive on how the Codex CLI agent loop works under the hood. Explains how Codex actually builds prompts, calls Responses API, caches, and compacts context. Reveals the exact mechanics that make a coding agent feel fast and stable, like relying on exact-prefix prompt caching to avoid quadratic slowdowns and using /responses/compact with encrypted carryover state to keep long sessions running inside the context window. Codex CLI is software that lets an AI model help change code on a computer. Codex is a harness that loops: user input → model inference → tool calls → observations → repeat, until an assistant message ends the turn. - The “agent loop” is the product So the model is the brain that writes text and tool requests, and the harness is the body that does the actions and keeps the conversation state tidy. In Codex CLI, the harness is the CLI app logic that builds the prompt, calls the Responses API, runs things like shell commands, captures the results, and repeats. It does the work by looping between model output and tool runs until a turn ends. The main trick is keeping prompts cache-friendly and compact enough for the context window, meaning the max tokens per inference call. Codex builds a Responses API request from system, developer, user, and assistant role messages, then streams back the sampled output text. If the model emits a function call like shell, Codex runs it and appends the output into the next request. Before the initial user message, Codex injects sandbox rules, environment context, and aggregated project guidance, with a default 32KiB scan cap. Because every iteration resends a longer JSON payload, naive looping is quadratic, so exact-prefix prompt caching keeps compute closer to linear on cache hits. - Codex is designed around stateless Responses API calls Codex avoids stateful shortcuts like previous_response_id to stay stateless for Zero Data Retention (ZDR), which makes tool or config changes a real cost. When tokens pile up, Codex calls /responses/compact and replaces the full history with a shorter item list plus encrypted carryover state instead of the older manual /compact summary.
@nrqa__ ·
OpenAI just open sourced the system they used to write 1 million lines of production code with zero human-written code. It's called Symphony. And it doesn't just write code. It manages the work itself. You point it at your Linear board. It reads the tickets. It spawns a Codex agent for each one. The agent codes, tests, opens a PR, attaches a walkthrough video as proof of work, and waits for review. You don't supervise agents anymore. You manage work. Here's what happened inside OpenAI: → A small team at OpenAI Frontier ran a 5 month experiment → Rule: zero human-written code. Everything had to come from Codex. → Result: 1 million+ lines of code shipped → ~1 billion tokens processed per day → 5 to 10 PRs per engineer per day → Some teams saw a 500% jump in landed PRs in 3 weeks One engineer reportedly shipped 3 significant code changes from a cabin with bad WiFi using only the Linear mobile app. The agents did the rest. Here's how it actually works: → Symphony is a daemon. It watches your issue tracker (Linear for now). → Every open ticket spawns an isolated Codex agent in its own sandbox. → The agent reads the harness. Codes. Runs tests. Opens a PR. → Attaches CI status, complexity analysis, code review feedback, and a walkthrough video. → You review at the PR level. Accept and it lands. Reject and the agent reworks. → New tickets show up. Symphony picks them up automatically. The shift is the part nobody is talking about: When humans stop reviewing every line, the perceived cost of a code change drops to near zero. Bugs that weren't worth fixing before are now worth fixing. Features that were too small to prioritize are now in scope. The economics of software flip. Here's the wild part: The team used Symphony to build Symphony. Once the basic loop worked, the agents wrote the rest of themselves. Internal product market fit was so strong inside OpenAI that other teams started adopting it organically. That's why it's public now. A few notes: → It's written in Elixir. ~1,000 lines. → OpenAI calls it a "low-key engineering preview." They are NOT maintaining it as a product. → It's a reference implementation. You're meant to point your own coding agent at the SPEC.md and have it build a version for your stack. → Works best in repos that have adopted "harness engineering" first. Tests, observability, agent-readable docs, the whole setup. The README literally says: "Tell your favorite coding agent to build Symphony in a programming language of your choice." That's the new pattern. Specs are the new code. Agents are the new compilers. Apache 2.0. 100% Open Source. repo: https://t.co/ZRAWuqOZLQ
@_vmlops ·
OPENAI CODEX IS SILENTLY DESTROYING YOUR SSD A github issue just revealed codex writes ~640 tb/year to a local sqlite feedback log → after 21 days of uptime, one user's ssd had written 37 tb → that's 640 full-drive writes/year on a 1 tb ssd → some consumer ssds are only rated for ~600 tbw total → 36,211 rows inserted every 15 seconds while retained count stays flat → 70% of it is TRACE-level noise: inotify events, websocket internals, locale file opens → root cause: global `Targets::new().with_default(Level::TRACE)` logs everything, always your drive could hit its entire warranted write endurance in under a year just from running codex the fix isn't disabling logs it's better defaults. raise thresholds, drop dependency noise, stop mirroring raw otel payloads to sqlite this isn't a rare edge case. multiple related issues show this has been happening across desktop and cli for months
@jonoringer ·
The single-threaded AI chat window isn't efficient enough.. i did a comparison between claude code and codex multithreaded architectures: Claude Code and OpenAI’s Codex handle multi-threading completely differently under the hood. Here is the technical breakdown and the playbook to combine both: - Architecture: Peer Networks vs. Hierarchical Queues • Claude Code uses a Collaborative Peer Network. It spins up independent Claude instances ("Teammates") that isolate execution using local Git Worktrees hidden inside your repository (.claude/worktree_A). Teammates use a direct mesh protocol via native SendMessage tools to talk to one another, negotiate code interfaces, and update a shared TaskList dynamically. • OpenAI Codex uses a Hierarchical Fork-Join Queue. The main thread encounters an objective, Forks into specialized, independent subagents inside secure Cloud-Sandboxed Virtual Containers (e.g., separating a Security Auditor from a Test Suite), executes them in parallel in the cloud, and then Joins and distills their outputs back into the main thread. - Managing the Multi-Threaded "Token Tax" .. Parallel processing introduces a massive token drain. • To Optimize Claude Code: Use a root CLAUDE.md file to enforce strict subdirectory routing rules, preventing redundant repository-wide context indexing. • To Optimize Codex: Use prompt constraints to force input truncation: "Spawn 3 subagents. Return only a 5-line markdown summary to the main thread and close the container immediately." The Dual-Agent Playbook.. trying this now. anybody else have experience here..? Maximize throughput by running them simultaneously as a heterogeneous cluster: • Step 1 (Parallel Auditing ➡️ Codex): Use Codex to concurrently sweep for security risks, performance bugs, and test gaps across your repo. It leverages cloud container speed and keeps your local workspace clean. • Step 2 (Deep Implementation ➡️ Claude Code): Feed Codex’s audit report to Claude Code. Command it to spawn teammates to update database models and refactor API routes concurrently via local Git worktrees. Because they share a task list, the agents will actively negotiate code boundaries mid-execution. • Step 3 (Background Verification Codex Automations): Offload long-running background tasks, regression checks, and CI/CD failure triages to Codex cloud containers while your local machine remains unburdened. See the full architecture diagram below..
@metalab ·
The State of Coding Agents for 2026. Our Engineering Director, Paul Straw, shares what he learned testing Claude Code, OpenAI Codex, and Gemini CLI in real production work. What helps, what doesn’t, and why benchmarks aren’t the whole story. 🔗 Full post + video in thread.
Watch video
@andrewsthoughts ·
We're leaving the "copilot" era and entering "AI orchestration." OpenAI's new Codex app joins the ranks and validates the category: parallel agents, reusable skills, and first-class scheduling. The missing primitives are about trust: → Workflow enforcement (eg a flavor of SDD) → Multi-agent verification https://t.co/a8pX0NnhS1
Best Tweets by Topic
