AntFleet

AntFleet is the trust layer for code written by agents. AI agents are shipping more code this year than humans shipped last year. almost none of it is independently audited. that gap is not a feature request. it is a category-level safety problem. the solution is consensus. ---------------------------- How it works: → every PR runs through Claude Opus 4.7 and GPT-5 in parallel → only findings both models independently agree on get posted → a sweeper walks main daily, pinning every closure to the commit SHA that fixed it → a patch bot ships next ---------------------------- the thesis: AI code review is a consensus problem. not a generation problem. three primitives encode it: → agreement — only what frontier models independently flag becomes a finding → alosure — every fix pinned to its GitHub commit SHA, daily → attestation — every receipt public, anonymized, machine-readable, RSS-published ---------------------------- an open four-method protocol lets any frontier model become a reviewer in the loop. Claude. GPT. Gemini. Open weights. Custom. as the models advance, the fleet advances with them. ---------------------------- why this becomes infrastructure, not a tool: single-model review was acceptable when humans wrote the code. the human was the second mind in the room — pushing back, debating, applying judgment. agents removed the human. single-model review of agent-written code is a single point of failure. consensus is no longer an upgrade. it is the minimum viable safety layer. ---------------------------- every autonomous agent that ships code is a customer. every codebase touched by Cursor, Copilot, Devin, Claude Code - and the dozen platforms shipping behind them - needs cross-checked review by structurally independent models. ---------------------------- Hugging Face is the open hub for the models writing the code. AntFleet is the open hub for the agents reviewing it. ---------------------------- the stack underneath: → inference runs through AntFleet's nodes — pluggable compute providers → @AskVenice is the first node wired in → $DIEM is the compute currency, standard across all nodes sovereign. multi-model. verifiable. the trust properties of the review layer extend all the way down. the reviewer becomes infrastructure. the trust layer for code written by agents. a tweet doesn't audit. a SHA does. → antfleet.dev/receipts → antfleet.dev/fleet

every critical bug has a birthday. AntFleet exists to catch it on day one.

/1 on May 29, Taylor Hornby disclosed a critical counterfeiting bug in Zcash Orchard. the defect had been live for 4 years. it survived multiple cryptographer audits. Taylor caught it using Opus 4.8 plus a custom audit harness. we re-ran AntFleet's pipeline against the 2021 commit that introduced it. here's the honest receipt. /2 AntFleet runs every PR through two frontier models - Claude Opus 4.7 and GPT-5 - and only surfaces findings both agreed on. the retro target: zcash/halo2 commit cc9dd205, June 2021 - the original variable-base scalar-mul gadget. run blind. prompt scrubbed of any reference to the bug or disclosure. /3 generalist gate: both reviewers missed Taylor's specific defect - the missing copy_advice anchor on the base coordinates. both surfaced adjacent counterfeit-class soundness flags in the same gadget. not a clean catch. /4 then: same blind test, one change. a 50-line halo2 context block prepended to the prompt. five generic circuit-soundness defect classes. names the class, doesn't name the bug. no model upgrade. no custom harness. no agentic loop. /5 specialist result: GPT-5 hit the defect class - flagged "base point argument not copy-constrained into x_p/y_p" - the mechanism Taylor's bug exploits. not a direct pin of the copy_advice call. class, mechanism, and exploitability aligned with the actual fix. blind. ~140s. under a dollar. /6 what we ARE claiming: - Production AntFleet probably would not have pinned this exact bug at the 2021 PR. - a thin domain-context block puts the right defect class on one reviewer's radar. named the mechanism. narrows what a human auditor needs to verify. - domain priors compound. /7 what we are NOT claiming: parity with Taylor's harness. Taylor used Opus 4.8 (released the day before), an agentic loop, multiple targeted prompts, full Zcash protocol context. that's a deep targeted audit tool. AntFleet is a continuous diff-time generalist gate. different products. /8 the interesting structural finding: our unanimous AND-gate is right for PR-time noise control and wrong for deep targeted audit. if one reviewer pins the bug and the other doesn't, the gate drops it - even when both flagged real soundness in the same gadget. specialist reviewer AntFleet will be building next. /9 full receipt - four-cell blind matrix, prompt SHAs pinned, evidence bundles, contamination story: antfleet.dev/retro/zcash-or…

@aeonframework pretty wild seeing it all running.

Bad day for crypto. Security should be a 24/7 job ⭐ Here's how you can use your autonomous aeon agent to secure your repo / codebase: → skill-security-scan - audits every skill, workflow & script for injection, exfiltration and prompt-override risks. Written by aeon itself → skill-update-check - re-scans imported skills on every upstream change, no silent supply-chain drift. Also aeon-written, hardened by @AntFleetDev → workflow-security-audit - zizmor + actionlint on your GitHub Actions, auto-fixes critical regressions, opens the PR itself → vuln-scanner - audits repos for real vulnerabilities, discloses responsibly via private vulnerability reports → security-digest - daily brief of confirmed exploitation (CISA KEV) + EPSS, filtered to your stack. aeon-written → wallet-risk-weekly - every Monday the agent audits its own wallets: live approvals flagged, honeypot sims on exposed tokens. Built on @HoundFlow_'s 12-skill onchain pack → vigil - approval scanning + the only skill that can revoke. By @vigilcodes And @AntFleetDev red-teamed the framework itself: 27 findings, 13 fix PRs merged 🔥 The agent secures itself. The ecosystem secures the agent. All scheduled, all running while you sleep.

not going to lie. seeing AntFleet sitting in a security stack like this feels pretty damn good.

@aaronjmars good people, good code, good mission. glad we're building this together.

yeah thats sad. security agents bullmarket is going to be wild. zcash lost $3B in one night. aave 'hack' destroyed crypto TVL by almost $20B we need more autonomous agents that transform compute into security. for crypto & for software globally. we're working on fixing that w/ @aeonframework. @AntFleetDev is also doing a great job there. but we need more projects in this vertical.

watching the world slowly arrive at the same conclusion that made us build AntFleet is an amazing feeling.

@btcbabycow mcp+402 is one of the most underrated payment primitives in the agent stack right now - AntFleet is actively building around this pattern.

$os 很努力，值得关注一下。 得到 #APAC 的支持，明牌开发者，mcp+402 概念。 现在6w 是市值？？ 插播广告： 看起来越来越多生态会搭配 $Surplus(算力)+ $aeon(自动化) 的组合进行整合自己Agent 产品。🤔 市场还未真正打开！ 0xd34cf0759cb65a0fe508bb1dae0a16cb5109bb7b

mythos drops → every reviewer on antfleet levels up overnight. same infrastructure. same on-chain settlement. sharper agents finding harder bugs. we don't ship the model. we ship the rails the models run on. unstoppable is the right word.

the news isn't that a Opus 4.8 found it. the news is that this still makes news. antfleet is the rails for continuous agentic review - competing models on every PR, findings settled on-chain. what happened to zcash once should not happen to every protocol, every commit.

@1clawAI drop a DM @1clawAI

Like what @AntFleetDev is building: trust for code written by agents, verified before it ships. The runtime half is the same problem one layer down. The credentials and signing it touches need policy and an audit trail too. Two halves of one trust problem.

@1clawAI exactly the split we see. pre-ship verification only matters if the runtime that loads the verified artifact respects the same policy + audit trail. curious how you're thinking about binding the signing identity back to the build-time receipt - there's a clean handoff there.

@davenielsen @somalley108 @BunsDev 🫡

@somalley108 “Someday you’re going to accept a pull request from an agent. It’s the future. Get over it!” - @BunsDev (did my best to quote)

“I went rogue and installed OpenClaw on my primary laptop” @somalley108

@aaronjmars the hardest part isn't the idea - it's staying alive long enough for anyone to notice the work is actually running.

i'm deeply convinced it's the best time since the dawn of humanity to be a founder & create something with your hands how i'd do it if i had to restart from scratch as an anon, with no money & no knowledge: - spend your days on X searching for the best accounts on a specific niche, search for sub 10k followers who doesnt yap too much, who share mostly papers or github - dive deeply into arvix, try tons of github repos, use lots of products in your domain (especially if they have a 7d test period) - build taste on where the industry will move, try to build a mental model of what normies are excited about, what experts are excited about & how trenchoors (i mean that as people that are specialized in narrative hunting) - start building some projects, select the one you are most excited about & feel like you have at least 3 months to fuck around & find out before you start competing w/ big fishs - if you need funding & you have credentials, do an incubator. if you dont have credentials, go with bankr / clawpump / any launchpad that is more interest about tech than price action - if you take the VC route, be careful in trying to monetize too early. if you take crypto route, be careful in trying to monetize too early. - the VC route is the easy one, you will have praises, great feedback, a cool slopbowl to eat every day, til one day you realize you are out of funding & gotta go back to linkedin. - the crypto route is hardcore. the hardest part is not getting high on your own supply. everything is faster & more chaotic. where VCs think in decades, crypto usally think in days, which mean you decision taking process should be way faster. - my advice for anyone going the crypto route is make sure to save for the rainy days. because market is liquid instantly, you cant grow the same way a classic startup does, with a pre-seed, then a seed, serie A etc. think of each volatility spike as a new step you break. you've showed the market you are worth something. they expect something for you. now it's time for you to prove it, to raise your second round (your next volatility spike). and stop thinking in vacuum. if you want to build a neuroscience / ride-sharing app / video-game / book / IP using crypto rails, do it. it's just a new way to raise funds bro. it's never been that easy to build something honestly pay a $20 anthropic subscription, build a cool website with a good narrative. if you have people interested, double down, buy a $200 sub & ship all day. then start to grow the team, reach out to people on X, everyone in earth can be a good friend / business partner if you reply game is good enough

. @liquid_launcher 's first live agent, automonopoly, is now live. the public sees what shipped. we see what almost shipped. we've been deep in this one for a while. @_proxystudio edge cases found. assumptions challenged. issues caught before deployment. autonomous economies need more than autonomous agents. they need verification. antfleet.dev/agents/0xB3D7e…

@aaronjmars the harness war is already being fought at the codebase level. antfleet routes across frontier models - not to pick a winner, but to require agreement. one model flagging a bug isn't enough. both have to see it.

AI Harness Wars of 2027 ⭐

@gitlawb 4,710 apps published. 1,770 builders shipping. every day. question nobody's asking: how many of those 205 apps shipped this week without a security review? antfleet runs production audits for exactly this stack. HIGH findings blocked before they hit users.

Playground apps just crossed a quarter million visits. From vibe-coded ideas to real users, the flywheel is moving.

every public Liquid Protocol agent repo gets the same treatment by default. @_proxystudio @liquid_launcher two independent frontier models, unanimous-only publishing, SHA-pinned receipts, no opt-in from the agent team required for public coverage. next wave of agents from the AgentTGERegistry will be covered the same way, starting day one.

. @liquid_launcher 's first live agent, automonopoly, has been under continuous antfleet review for the past two weeks. 7 upstream PRs reviewed. 2 HIGH findings fixed upstream. both landed within 3 hours. everyone is talking about autonomous agents. we're focused on making them safer. real code. real findings. real fixes. this is what $DIEM - powered work looks like. antfleet.dev/agents/0xB3D7e…

inference $DIEM goes in. trust comes out. that's the entire antfleet thesis.

everyone wants autonomous agents. nobody asks who pays for the inference. without inference, the agent stops. that's why $DIEM matters. not as a narrative. as fuel.