Amart (LOOP)⚡️

@bwarrn super cool, going to try it out with my agent

I'm so excited to officially share what we've been cooking up at Mesa: the most powerful filesystem ever built for AI agents. The dirty secret of every "production" AI agent today: the filesystem is held together with duct tape. Teams are stitching together S3, GitHub, sandbox-local disks, and homegrown diff logic to give their agents something resembling persistent, versioned storage. None of it works. S3 isn't designed for parallel agents - concurrent agent writes silently overwrite each other. GitHub has the semantics but rate-limits you into the ground at agent scale and doesn't give you filesystem ergonomics. Sandbox disks vanish the moment the container dies. And your agents don't want to git clone and git push anyway. 𝗧𝗵𝗲𝘆 𝘄𝗮𝗻𝘁 𝘁𝗼 𝗿𝗲𝗮𝗱 𝗮𝗻𝗱 𝘄𝗿𝗶𝘁𝗲 𝗳𝗶𝗹𝗲𝘀. 𝗟𝗶𝗸𝗲 𝗲𝘃𝗲𝗿𝘆 𝗽𝗿𝗼𝗴𝗿𝗮𝗺 𝗲𝘃𝗲𝗿 𝘄𝗿𝗶𝘁𝘁𝗲𝗻. So we built the missing layer. Mesa is a durable, POSIX-compatible filesystem with version control built in. Branches, diffs, history, rollback, access control — every primitive a codebase has, for any file type, at agent scale. You mount it. Your agent uses it like a normal filesystem. We handle the rest. Private beta is live. Link in comments.

@businessbarista @tenex_labs feels like it could be a “claude do this” task, no? maybe i’m missing something

I may be crazy, but I built a 20-level excel game to find a Finance savant to join our company. The game is called "Bug Hunt," and any excel junkie interested in becoming the leader of our Finance function at @tenex_labs can play. If you complete all 20 levels, you are accelerated to a final round interview with my cofounder & me. Here's how it works: 1) Open the model. It's a live workbook in your browser with the "finished" financials of a fictitious SaaS company. 2) Mark every bug. Click any cell, write one line of reasoning. Submit when you're sure. There are 20 total. 3) Climb the tiers. Each correct catch unlocks the next. The final three are veteran CFO-level. 4) Hit level 20 & auto-move to a final round interview. Play the game: web-production-42101.up.railway.app P.S. you can still apply to be our Senior Director of Strategic Finance (application below) the normal way, it's just a little less fun & you don't get an auto-invite to final round.

@trillhause_ respectfully disagree my favorite tools are ai that scrape a companies date and builds a system custom based on their current system

This is going to be a tarpit idea. It’s good in theory, but impossible to pull off unless it’s an internal company effort by a tyrant like CEO. An external company will never be able to build a software that results in a company brain. It’s mostly because no tool will have perfect adoption from all employees and data will always be fragmented across new systems. Chaotic systems are very hard to capture. It’s impossible to perfectly extract data from all sources as companies evolve and introduces new data sources. You will spend all the time keeping track of the data instead of doing actual work. This is same trap that the second brain productivity folks fall for.

@ml_angelopoulos a couple months?

Why GPT-5.5 is lower than Claude? The answer is simple: Code Arena currently only supports frontend/web development tasks, where GPT-5.5 is weakest. Full-stack app development and GitHub integration will land in a couple months. Next time we'll be clearer that this leaderboard shows React/FE only, until we ship full-stack apps etc. Thanks for the feedback!

Simply Lovely !

I do AI implementation at startups + large companies one is an M&A firm, the 50+ year olds absolutely HATE ai like really hate it I show them they can turn the annoying manual spreadsheet process they've been doing for 10 years into to 1 skill /skill + enter then it's done they change their minds pretty quickly

Actual anti-AI energy is the newest instance of @robkhenderson’s “luxury beliefs” — a small fringe of elites playing at radical politics. For normal people, AI is a useful app on their phone, and they’ll just use it more and more.

.@steipete GPT 5.5 with openclaw is simply magic incredible work 🦞

@vincent_koc @douglasfugazi some of these questions really make me laugh

@douglasfugazi What "gateway issue"?

The great purge release. A monster load of issues we could not see are finally in after the culling of slop and noise. Enjoy your new freshly minted lobster.

@thsottiaux i was at 2% bless up

Don't just reset Codex rate limits for fun, it costs money. Don't just reset Codex rate limits for fun, it costs money. ... but the vibes are good ... I have reset Codex rate limits for ALL paid plans to celebrate a good week and allow everyone to build more with GPT-5.5. Enjoy

@steipete @allanmelsen HAHA this is a great response obviously op has good intentions but this is great

@allanmelsen Have some patience.

the crawl army so agents can read it all.

@fxnction @openclaw use github, make sure they work in separate project folders, then use a CEO agent to manage them if working in the same project, have them take the “lock” so it shows the other agents which code they’re working on

Anyone else running more than 3 @openclaw agents? How do you keep them from stepping on each other?

@Voxyz_ai yeah significantly better, can also run claude via ACP sessions which works great

feels like openclaw is really back🦞. spent a lot of time yesterday just chatting with mine. would recommend gpt 5.5 + fast mode. speed is great too. you can try deleting every "allowed / permitted / when appropriate" from your soul. permission language reads as optional, and the model doesn't always reach for it. "you have dry wit" holds up. "humor is allowed" lands weaker. one is identity, the other is a switch. identity stays on by itself. switches need the model to actively flip them, and the model doesn't really do that.

@hunvreus ai will tell them not to hire you for this tweet

Talking to smarter folks than me, I'm convinced many of the AI folks in my timeline are full of shit. Nobody is "running 20 agents over night" and building stuff for actual users. Maybe some are building internal tools or disposable software. Maybe. But building software people like using? That doesn't get hacked on day one or blow up after the 3rd user? Nope. I don't even understand what that's supposed to look like. Do you work out a 57 pages document that perfectly describes what you want to build and then summon 14 agents and have them run wild for 6 hours? And what comes out on the other end isn't a broken pile of shit? Nope. Not buying it. PS: it may also be that I have an IQ of 82 and can't figure it out.

@KingBootoshi very helpful when it becomes a system or skill you can just call and have it go for 4 hours and build something flawless anything you’ve found particularly helpful or hurtful in your prompting/ skill making

when I set up an orchestration flow using Claude Code to execute a very direct PRD with Codex agents, my runs take 3-4 hours to fully complete so I often have this running overnight because I've automated my entire prompt workflow I usually do first off you have to understand that creating and using agents are a form of alchemy. one wrong word/prompt/step and your output went from a God tier one shot to being absolute dog shit anyways my workflow i have agents do is research, scoping out the codebase, doing the actual implementation, then review. research/scoping code prevents hallucinations and induce a genuine understanding of the design and patterns set review agents will flag things that the PRD requested which got missed, or flawed bugs then the {manager agent} will handle spawning in these specialized sub agents per task. the main context window as a result of using subagents remain small, which lets it do a lot of managing to keep the entire workflow focused i'll often have a couple of these workflows running at the same time, because they LITERALLY take hours. so I just relax after setting off an army as a result, workflows like this mean I have one agent handling up to 16+ subagents throughout the whole sub process i also combine this with hard guards to enforce good design like eslint & custom rules, strict design patterns and a lot of foundational templates i've built for myself the last year a lot of what i do end up becoming skills so i'll literally /skill combo my way through a prompt to chain up a series of very detailed actions at the end of the full process, THEN i review the full PR and it's really easy for me to skim through the code this is absolutely not vibe coding, and instead taking engineering workflows i've fundamentally learned and automating the process with detailed agents there are levels to this shit. the smallest discovery could 2x the quality of your workflow. sometimes 10x, even more. you need to be open and keep exploring!

@doodlestein there it is 🔥

@LoopOnChain All my other tooling is open source, see: agent-flywheel.com/tldr

I finally got around to making a skill a lot of people have been asking me for: jeffreys-skills.md/skills/simplif… It basically helps to "de-slopify" and refactor code that's been written by agents, looking for ways to simplify and reduce the amount of code without changing the behavior. The difference between this and other skills or prompts in the same spirit is the lengths this one goes to in order to prevent the process from going off the rails and introducing bugs or security problems. It's a whole elaborate system spanning 98 files and one full megabyte of reference files, scripts, and subagents (see pic). You can run it over and over again and it will autonomously identify good opportunities for accretive simplification and do everything needed to implement the changes and prove that they didn't change the outputs. GPT-5.5 can explain better than I can how it does all that and what makes it so compelling and useful: --- The strongest thing about this skill is that it treats refactoring as a proof obligation. A normal “clean this up” prompt invites the model to follow taste. It sees repetition, long files, wrapper functions, stale types, try/catch clutter, _v2 files, and it starts cutting. Sometimes that works. Sometimes it silently changes error semantics, loses a side effect, removes a lifecycle hook, or deletes a file that looked unused but was actually the intended implementation path. This skill changes the frame. A simplification claim becomes: “this smaller program is observably equivalent to the larger one.” Then it makes the agent prove that claim. It starts with a baseline: tests, golden outputs, LOC, warnings, complexity. It maps duplication instead of eyeballing it. It classifies clones, because exact copy-paste, parametric duplication, semantic similarity, and accidental rhymes are completely different things. It scores each candidate by expected LOC saved, confidence, and risk. Low-score candidates get rejected and logged, which is important because future agents otherwise rediscover the same bad idea forever. The isomorphism card is the key move. Before editing, the agent has to answer boring but lethal questions: same ordering, same errors, same logs, same metrics, same side effects, same async cancellation behavior, same React hook identity, same serialization, same resource lifecycle. Those rows catch the kind of bugs that compilers and ordinary tests miss. Then the edit discipline is deliberately narrow: one lever per commit, no rewrites, no sed, no drive-by fixes, no deletion without explicit permission. Afterward, it verifies behavior again and records the result in a ledger. If the refactor did not actually preserve behavior, it does not get to call itself a refactor. What I like about it is that it matches the real failure modes of agent-written code. AI code tends to accumulate plausible junk: defensive branches for impossible inputs, duplicated wrappers, too many optional parameters, orphaned “improved” files, shallow happy-path tests, stale types, and comments that are really leftover task plans. The skill has a whole pathology catalog for those patterns, plus scripts and subagent roles to find them systematically. So the compelling part is not “make the code prettier.” The compelling part is leverage with brakes. You can send very strong models into messy codebases and ask them to reduce complexity aggressively, while forcing them to preserve the contract that matters: observable behavior. That is the difference between a refactor you hope is safe and a refactor you can audit.

@pashmerepat Seems like we need to set codex as the default embedded harness

@pashmerepat Do we need to tell our agents to start using codex as the default harness or is this automatic when using the open ai codex oauth?

I've embarked on a new sprint. My mission is to make OpenAI models feel magical in OpenClaw in the next few weeks. Diving in today, I noticed a bug. When you configured OpenClaw to use the Codex harness with OpenAI models, auth was broken, and the system was silently falling back to the Pi harness. So nobody knew it was broken. Two PRs later (fix the auth bridge, stop the silent fallback), the Codex harness actually works. And the difference is night and day (pic related). Before: the agent didn't feel magical or proactive. It did the exact same shallow loop every heartbeat. Read the heartbeat file, check Discord, see nothing, say HEARTBEAT_OK. It ignored the rest of its instructions. Sometimes it would even reason about doing work and then just... not issue the tool calls. After: full agent loops. It reads its workspace context, interprets the entire checklist, inspects the repo, makes real edits, tries to verify them, and gives honest status reports when things are blocked. Later heartbeats show continuity, it doesn't repeat work, it picks up where it left off. I didn't change any prompting or scaffolding. Just swapped in the codex harness for pi. Lesson here is use the codex harness if you're building with OAI models. A lot more to do but this is a strong start.

Internal productivity of a company is now the new moat Ship faster, yell louder You win

@_MaxBlade will be a good week

I am so confident this week Peter + Open Ai team will bring openclaw back to life. Opus is the only model that works right now, but pay per use ruins the entire experience. Instead of building and exploring your constantly left thinking about how much this is going to cost you. It’s too important to let die.