AICoder

AI Coding Live News

First-hand AI coding updates from 𝕏/Twitter frontier labs and researchers, distilled by AI with entity benchmarks.

Explore live on 𝕏 ↗

AI Coding Live News

N
NotionNotionHQ·
👀0
🔥88
View on 𝕏 ↗

Notion: MCP Shared Brain Lets ChatGPT, Claude, Muse and Bot Share Memories and Skills, With Official Agent Setup Prompts

On November 4 NotionHQ announced that via Notion MCP, ChatGPT, Claude, Muse and Bot can read and write the same Memories and Skills—one shared brain across agents. Official setup prompts (v0.0.6, Oct 30), help docs and a Nov 3 MCP use-cases blog corroborate the product push.

Notion: MCP Shared Brain Lets ChatGPT, Claude, Muse and Bot Share Memories and Skills, With Official Agent Setup Prompts

Key Takeaways

  • NotionHQ post (Nov 4, 19:14 UTC) had about 3,271 likes, 225 reposts, 4,560 bookmarks and ~589K views when fetched
  • Product claim: ChatGPT / Claude / Muse / Bot share Memories and Skills in Notion so switching agents does not reset context
  • Official setup prompt v0.0.6 (2026-10-30): walks agents through an Agent Operating Guide and custom-instruction paste for ChatGPT / Claude
  • Help center: Notion MCP works with Claude, Cursor, ChatGPT Pro and more; Enterprise can approve a whitelist and govern connections
  • Vendor blog (Nov 3): five team stories of MCP workflows as reusable SOPs; Skills via MCP are listed as coming soon
Read full story
Linked Models & Agents:
🤖 ChatGPT🤖 Claude🤖 Muse🤖 Notion Bot🌐 Notion
#Notion MCP#shared memory#Skills#agent setup#ChatGPT#Claude#Tooling
G
Guillermo Rauchrauchg·
👀3
🔥82
View on 𝕏 ↗

X Discussion: Vercel CEO Guillermo Rauch Says AI Will Rewrite All Human-Made Software, Sparking Debate on Formal Verification and Rewrite Risk

On November 5 Guillermo Rauch posted on X that most human-made software was slow, insecure and under-specified, and that AI will rewrite it all, many times, then formally verify, load test and harden it: "the true birth of software engineering." The post drew 3,000+ likes, 270+ reposts and nearly 100K views. Supporters cite from-scratch AI rewrites; critics warn that missing specs and human review still matter. This is a personal opinion, not a Vercel announcement.

Key Takeaways

  • Rauch's post (Nov 5, 02:22 UTC) had about 3,060 likes, 272 reposts, 681 bookmarks and ~99K views when fetched
  • Core claim: AI will rewrite human software "many times", formally verifying, load testing and hardening it, and replacing memory-unsafe foundations
  • An hour earlier he shared Vercel Security Review figures: 6,817 vulnerabilities found in 24 hours, 580 critical/high, 1,411 already patched by Agent (vendor-reported)
  • Pushback: rewriting under-specified code can just ship the same wrong behavior faster, and formal verification moves bugs up an abstraction level rather than eliminating them
  • This is an opinion thread, not an established fact or official roadmap
Read full story
Linked Models & Agents:
🤖 Vercel Agent🌐 Vercel
#X discussion#AI rewrite#formal verification#software security#Guillermo Rauch#Vercel
J
Justin Ubertijuberti·
👀6
🔥86
View on 𝕏 ↗

OpenAI Realtime Lead: GPT-Live Outbound Telephony API Hits GA via Any SIP Trunk, with HD Voice and DTMF

On November 4, Justin Uberti, Head of Realtime AI at OpenAI, said the GPT-Live outbound telephony API is generally available: Live voice agents can dial through any SIP trunk such as Telnyx or Twilio, with HD Voice and bidirectional DTMF. Official docs describe Live API + SIP transport; Greg Brockman amplified the note.

OpenAI Realtime Lead: GPT-Live Outbound Telephony API Hits GA via Any SIP Trunk, with HD Voice and DTMF

Key Takeaways

  • Justin Uberti (Head of Realtime AI, OpenAI): GPT-Live outbound telephony API is GA via any SIP trunk
  • Official docs: Live API POST /v1/live/sessions with transport.type=sip; HD Voice and bidirectional DTMF
  • Trunks need TLS (sips:) signaling and SDES-SRTP; examples cite Telnyx and Twilio
  • Documented limits: ~3 minutes ringing and ~2 hours connected call duration
  • Amplified by Greg Brockman; frame as a first-hand employee note, not an @OpenAI press post
Read full story
Linked Models & Agents:
⚡ OpenAI GPT Live 1🌐 OpenAI
#OpenAI#GPT-Live#outbound telephony#SIP#DTMF#HD Voice#voice agents
G
GitHub ChangelogGHchangelog·
👀4
🔥81
View on 𝕏 ↗

GitHub Copilot Comes to JetBrains Air: Run Coding Tasks Locally in the IDE or in Isolated Cloud Environments via ACP, Billed to Your Copilot Subscription

On November 3 GitHub announced on its official changelog that GitHub Copilot is now available in JetBrains Air. The integration is built on the Agent Client Protocol (ACP): developers pick Copilot in Air's agent picker, delegate coding tasks, track several sessions in parallel and review changes next to their code. Tasks can run locally in a JetBrains IDE or in isolated cloud environments through Air Teams, and they draw on the Copilot subscription rather than JetBrains AI credits.

GitHub Copilot Comes to JetBrains Air: Run Coding Tasks Locally in the IDE or in Isolated Cloud Environments via ACP, Billed to Your Copilot Subscription

Key Takeaways

  • Copilot now appears in JetBrains Air's agent picker and works with an existing Copilot account
  • The integration uses the Agent Client Protocol (ACP) and can be used alongside the existing GitHub Copilot plugin
  • Developers can track multiple agent sessions in parallel and review changes next to their code using project context and IDE tools
  • Cloud tasks run through Air Teams, each in a Docker container on a fresh virtual machine, with traffic through a proxy and no state kept between tasks
  • Cloud tasks use the Copilot plan rather than JetBrains AI credits, but Air analytics do not include Copilot usage and admins cannot set Copilot spending limits inside Air
Read full story
Linked Models & Agents:
🤖 GitHub Copilot🤖 JetBrains Air🌐 GitHub🌐 Microsoft🌐 JetBrains
#GitHub Copilot#JetBrains Air#Agent Client Protocol#ACP#cloud agents#Air Teams#IDE
ADSponsored
C
ClaudeDevsClaudeDevs·
👀5
🔥82
View on 𝕏 ↗

Claude Code Background Subagent Status Lines Now Written by Haiku 5.5: About 99% Cheaper per Update, About 6% More Prompts per Limit

On November 4 ClaudeDevs, Anthropic's official developer account, said the status line that a background subagent in Claude Code refreshes every 30 seconds used to re-read its whole conversation each time, and is now written by Haiku 5.5 from short notes, so each update costs about 99% less. Anthropic says sessions with background subagents used about 5% less of the limit, giving about 6% more prompts on the same limit. It is live for most Pro, Max and Ultra users and everyone on Team and Enterprise, on v2.1.316 or later.

Key Takeaways

  • A background subagent refreshes its status line every 30 seconds, and each refresh used to re-read the subagent's whole conversation
  • Haiku 5.5 now writes the status line from short notes, which Anthropic says makes each update about 99% cheaper
  • The subagent itself keeps its own model; only the status line is written by Haiku 5.5
  • Anthropic says sessions with background subagents used about 5% less of the limit, so the same limit gives about 6% more prompts
  • It is live for most Pro, Max and Ultra users and everyone on Team and Enterprise; update to v2.1.316 or later
Read full story
Linked Models & Agents:
⚡ Claude Haiku 5.5🤖 Claude Code🌐 Anthropic
#Anthropic#Claude Code#Subagents#Haiku 5.5#usage limits#v2.1.316
C
Cat Wu_catwu·
👀10
🔥84
View on 𝕏 ↗

Cat Wu, Formerly on Claude Code at Anthropic, Announces She Is Joining OpenAI to Lead Product Management for Core Products

Cat Wu (@_catwu) announced on X on November 5 that she is joining OpenAI to lead product management for Core Products; her X bio says she previously worked on Claude Code at Anthropic. She said she will work with Tibo (@thsottiaux) and the PM team to build agents that have users' and teams' backs and keep them in control. The post drew about 6,900 likes, and Tibo posted a welcome. OpenAI has not published an official announcement.

Key Takeaways

  • Cat Wu announced on X herself that she is joining OpenAI to lead product management for Core Products
  • Her X bio says she previously worked on Claude Code at Anthropic, and before that at Dagster and Scale AI
  • She named Tibo (@thsottiaux, bio: Codex & ChatGPT @OpenAI) as a collaborator, with a focus on agents that keep users in control
  • Tibo welcomed her with a "Welcome Cat" post; her announcement drew about 6,900 likes and nearly 900 replies
  • OpenAI has not issued an official announcement, and which products Core Products covers was not specified
Read full story
Linked Models & Agents:
🤖 Codex🤖 Claude Code🌐 OpenAI🌐 Anthropic
#OpenAI#Anthropic#Cat Wu#Core Products#Codex#Claude Code#talent move
T
Thorthorwebdev·
👀6
🔥80
View on 𝕏 ↗

Gemini 3.8 TTS Adds Voice Remixing: Edit Prebuilt or Designed Voices With Natural-Language Instructions Into New Reusable Voice IDs

The official Gemini API changelog released Voice remixing on November 4. Starting from a prebuilt voice, an Extended Voice Library voice or a Voice Design persona, developers describe changes such as accent, pitch, pacing or age in natural language, and POST /v1beta/voices returns a new standalone voice_ ID usable with Gemini 3.8 Flash TTS and Flash-Lite TTS. It is also available interactively in Google AI Studio.

Key Takeaways

  • New Voice remixing feature: apply relative natural-language edits to an existing voice to create a new standalone voice
  • Uses POST /v1beta/voices (type=prompted, store=true, plus a new prompted.base_voice field) and returns a voice_ ID with a 24kHz WAV preview
  • Base voices can be any of 30 prebuilt voices, an Extended Voice Library voice, or a stored custom voice_ voice in the project
  • The resulting voice_ ID works with gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts and can itself be remixed again
  • The original voice is never modified, and deleting it does not affect voices remixed from it
Read full story
Linked Models & Agents:
⚡ Gemini 3.8 Flash TTS⚡ Gemini 3.8 Flash-Lite TTS🌐 Google🌐 Google DeepMind
#Gemini API#TTS#Voice remixing#Voice Design#Google AI Studio#speech generation
G
GitHub Changelog@GHchangelog·
👀7
🔥82
View on 𝕏 ↗

GitHub Copilot Comes to Xcode's Built-In Coding Assistant in Public Preview With Xcode 27.3 Beta, Supported on All Copilot Plans

GitHub's official changelog says Copilot now works directly in Xcode's built-in coding assistant, in public preview starting with Xcode 27.3 beta. It needs macOS Tahoe 26.6 or later, is set up under Xcode > Settings > Intelligence, and supports all Copilot plans; the existing Copilot for Xcode extension stays available.

Key Takeaways

  • Copilot joins Xcode's native coding assistant in public preview, starting with Xcode 27.3 beta
  • Requires macOS Tahoe 26.6 or later
  • Set it up under Xcode > Settings > Intelligence; all Copilot plans are supported
  • The existing GitHub Copilot for Xcode extension remains a separate option
  • Apple's Xcode 27.3 beta 1 release notes list it under Coding Intelligence
Read full story
Linked Models & Agents:
🤖 GitHub Copilot🌐 GitHub🌐 Apple🌐 Microsoft
#GitHub Copilot#Xcode#Xcode 27.3#Apple#coding assistant#public preview
I
Ivan Leo@ivanleomk·
👀11
🔥86
View on 𝕏 ↗

Google Antigravity Launches Remote Agents: Agent Sessions Run in a Google-Managed Cloud Environment and Keep Working After You Close Your Laptop, Rolling Out to AI Pro and Ultra

Ivan Leo, who works on developer experience at Google DeepMind, shared the official docs for Antigravity's new Remote Agents feature on X on November 4. Per the docs, Remote Agents run agent sessions and tools in an isolated, Google-managed remote environment, so tasks keep going when you close your laptop or browser or switch between desktop and mobile. Sessions start from the Antigravity App, the Antigravity CLI (1.4.12+, agy --remote-agent) or the web at antigravity.google.com, and remote projects can attach GitHub repositories, Google Drive files and uploads. It is rolling out gradually to Google AI Pro and Ultra subscribers on personal accounts; it is not available on the free plan, AI Plus or business accounts.

Key Takeaways

  • Remote Agents run in an isolated Google-managed environment with no local machine required; Remote Control instead connects to your own computer or server, and the web sidebar switches between them.
  • Three entry points: a New Remote Project in the Antigravity App, agy --remote-agent in CLI 1.4.12+, or antigravity.google.com in any browser, including mobile.
  • Remote projects can attach GitHub repositories (cloned automatically), Google Drive files and uploads (8 MiB per file, 100 MiB per project, 2 GiB per account).
  • CLI remote sessions default to the always-proceed (turbo) permission preset, so commands run without approval; locally installed MCP servers, plugins and skills are not available remotely yet.
  • Only for Google AI Pro and Ultra subscribers on personal accounts, rolling out gradually; not on the free plan, AI Plus or Workspace/Cloud business accounts.
Read full story
Linked Models & Agents:
🤖 Google Antigravity🤖 Antigravity CLI🌐 Google🌐 Google DeepMind
#Antigravity#Remote Agents#cloud agents#coding agent#Google
A
Artificial AnalysisArtificialAnlys·
👀15
🔥82
View on 𝕏 ↗

Artificial Analysis Makes Time per Task Its Headline Performance Metric: GPT-6 Astra Finishes a Max-Effort Task in About 17 Minutes, Ahead of Opus 5.5 and Gemini 4 Carbon

Independent benchmarker Artificial Analysis announced on X on November 4 (17:34 UTC) a new version of its API performance benchmarking: the headline metric moves from output speed to Time per Task, the time to finish one Intelligence Index task (output tokens per task divided by end-to-end output speed). It covers about 1,000 serverless endpoints, measures speed from sending the request to the final token, and adds multi-turn agentic coding and knowledge-work workloads. Per its data, GPT-6 Astra at max effort runs at 51 tokens/s yet finishes a task in about 17 minutes, against about 26 for Claude Opus 5.5 (max) and about 29 for Gemini 4 Carbon (high).

Key Takeaways

  • Time per Task equals output tokens per Intelligence Index task divided by end-to-end output speed, so a model that is fast per token but verbose can finish later than a slower model that uses fewer tokens.
  • End-to-end output speed runs from sending the request to the final token, including the wait before the first token (prefill and hidden reasoning); output tokens now come from each provider's API-reported counts instead of the o200k_base tokenizer.
  • GPT-6 Astra (max) runs at 51 tokens/s with about 50k output tokens per task and takes about 17 minutes; Claude Opus 5.5 (max, 111 tokens/s) takes about 26 minutes and Gemini 4 Carbon (high, 71 tokens/s) about 29, using roughly 120k to 170k tokens per task.
  • At each lab's API default effort (medium for all three), GPT-6 Astra, GPT-6.1 Sol and Opus 5.5 cut Time per Task by about 3 to 5x and Cost per Task by about 2 to 4x, scoring about 5 to 11 points lower; Opus 5.5 drops from about 26 minutes and $10.04 to about 6.8 minutes and $2.45.
  • The official methodology page lists this as version 3.0.0 (October 30), with runs now every 6 hours and a new endpoint error rate metric.
Read full story
Linked Models & Agents:
⚡ GPT-6 Astra⚡ GPT-6.1 Sol⚡ Claude Opus 5.5⚡ Gemini 4 Carbon🌐 Artificial Analysis🌐 OpenAI🌐 Anthropic🌐 Google
#Artificial Analysis#Time per Task#Benchmark#API performance#Agentic#Latency
ADSponsored
R
Ricursive IntelligenceRicursiveAI·
👀19
🔥84
View on 𝕏 ↗

Ricursive Open-Sources Ripple-65B Preview: A 65B Dense Model Pretrained From Scratch, TerminalBench 2.1 80.5, Apache 2.0

Ricursive Intelligence released its first model, Ripple-65B Preview, on November 4 (18:24 UTC), open-sourcing the instruction-tuned model (BF16 and NVFP4) and a base model on Hugging Face under Apache 2.0. The official model card describes a 65B dense Transformer pretrained on 12T tokens with up to 1M context, and self-reports TerminalBench 2.1 80.5, SWE-bench Pro v2.0 Hard 79.1 and DeepSWE 1.1 44.3, ahead of much larger US open-weight models on agentic coding. It fits on local hardware such as the DGX Spark. CEO Anna Goldie says it cost about a tenth of the prior comparable state of the art.

Ricursive Open-Sources Ripple-65B Preview: A 65B Dense Model Pretrained From Scratch, TerminalBench 2.1 80.5, Apache 2.0

Key Takeaways

  • Ripple-65B Preview is a 65B dense model with a thin-and-deep shape (80 layers, d_model 5120), moving from an Olmo 3 baseline to a Gemma 3N-like architecture.
  • It was pretrained on 12T tokens at 8K sequence length, then went through mid-training, long-context training, SFT and mixed-domain RL, with up to 1M context.
  • Self-reported TerminalBench 2.1 is 80.5, above Reflection Beam 501B (80.1) and Muse Spark Open 816B (76.4), but SWE-bench Verified 76.5 and DeepSWE 1.1 44.3 show it does not lead everywhere.
  • Ricursive says the core team grew from one to five people and code compute efficiency improved 16x versus the Olmo 3 ladder; it reaches about 250 tok/s on 2x B200.
  • Hugging Face hosts BF16, NVFP4, base and DFlash2 speculative-decoding checkpoints, all under Apache 2.0.
Read full story
Linked Models & Agents:
⚡ Ripple-65B Preview⚡ Reflection Beam 501B⚡ Muse Spark Open⚡ Nemotron 3 Ultra🌐 Ricursive Intelligence🌐 NVIDIA
#Ripple-65B#Ricursive#open weights#Apache 2.0#agentic coding#dense model#Local
C
Cognition@cognition·
👀26
🔥85
View on 𝕏 ↗

Cognition Launches SWE-2 Flash: A Fast Coding Model Built on Inkling-Small, 200 tok/s at One-Fifth the Price of SWE-2

Cognition released SWE-2 Flash on November 4 (19:55 UTC). Its official blog says the model is built on Thinking Machines' Inkling-Small (276B total, 12B active parameters) and, for the first time, adds a large mid-training stage before reinforcement learning, lifting FrontierCode 2.0 from 3.2% to 24.0% and Terminal-Bench 4.0 from 0.9% to 20.0%. Cognition says it beats every American open-weight model. It generates about 200 tokens per second at one-fifth the price and 3x the speed of SWE-2, and is free for Pro, Max and Teams users in Devin Desktop and CLI until the end of November. Weights are not public; a secure sharing program for partners and researchers is planned.

Cognition Launches SWE-2 Flash: A Fast Coding Model Built on Inkling-Small, 200 tok/s at One-Fifth the Price of SWE-2

Key Takeaways

  • SWE-2 Flash is built on Thinking Machines' Inkling-Small, with 276B total and 12B active parameters, about one-tenth the size of SWE-2 (2.8T-A104B).
  • Training adds a three-stage mid-training run on a few trillion tokens of scientific articles, blogs, code and other sources, followed by large-scale RL with higher-quality and more diverse tasks.
  • Official scores: FrontierCode 2.0 24.0 (SWE-2 scores 23.8), DeepSWE 1.1 60.1, Terminal-Bench 2.1 89.2, SWE-Bench Pro v2 Hard 89.9 and Terminal-Bench 4.0 20.0.
  • It runs at about 200 tokens per second versus about 60 for SWE-2, at one-fifth the price, and is free for Pro, Max and Teams users in Devin Desktop and CLI through November.
  • Citing cyber risk, the weights are not openly released; Cognition says details of a secure weight-access program will follow in the coming weeks.
Read full story
Linked Models & Agents:
⚡ SWE-2 Flash⚡ SWE-2⚡ Inkling-Small🤖 Devin🌐 Cognition🌐 Thinking Machines
#SWE-2 Flash#Cognition#Devin#coding model#Inkling-Small#FrontierCode
D
David Stout@Davidstout·
👀10
🔥82
View on 𝕏 ↗

webAI Open-Sources TwIL-8B: An 8.8B Reasoning Model With LiveCodeBench 81.6 and MATH-500 98.2, Under a Non-Commercial License

webAI CEO David Stout announced on X on November 4 (17:37 UTC) that the company is open-sourcing TwIL-8B. The official Hugging Face model card says it is post-trained from IBM Granite 4.2 8B with full-parameter SFT, GRPO/DAPO reinforcement learning and a length-aware RL stage, and ships in balanced and verbose variants. Self-reported scores include LiveCodeBench 81.6 (verbose), MATH-500 98.2 and IFBench 74.4, about 93% and 92% of the best frontier score, while GPQA Diamond is only 70.2. A 4.98 GiB Q4_K_M GGUF runs locally; the license is the webAI Non-Commercial License.

webAI Open-Sources TwIL-8B: An 8.8B Reasoning Model With LiveCodeBench 81.6 and MATH-500 98.2, Under a Non-Commercial License

Key Takeaways

  • TwIL-8B has 8.79B parameters, is built on IBM's ibm-granite/granite-4.2-8b, has a 131,072-token context window and is English-only, text to text.
  • Self-reported scores: MATH-500 98.2, AIME 2025 88.3, LiveCodeBench 81.6 (verbose) and IFBench 74.4 (verbose); GPQA Diamond at 70.2 trails Qwen3.5-9B (81.7) clearly.
  • The balanced variant averages about 19.5k output tokens per AIME 2026 problem, about 38% fewer than Qwen3.5-9B's 31.2k, but still roughly 4x GPT-5.5.
  • All scores are webAI internal evaluations; competitor figures are quoted from public sources and were not re-run. SWE-bench, tool use and agentic tasks were not evaluated.
  • The license is the webAI Non-Commercial License 1.0, so it cannot be used commercially as is; the Q4_K_M GGUF is about 4.98 GiB and runs with llama.cpp.
Read full story
Linked Models & Agents:
⚡ TwIL-8B⚡ Granite 4.2 8B⚡ Qwen3.5-9B⚡ Gemma 4 31B🌐 webAI🌐 IBM
#TwIL-8B#webAI#open weights#small model#Reasoning#GGUF#Local
C
Charlie Holtz@charlieholtz·
👀22
🔥85
View on 𝕏 ↗

Conductor Is Joining Meta: The Multi-Agent Coding Tool Team Moves to MSL, Conductor Cloud Shuts Down on November 20

Charlie Holtz of Conductor announced on X on November 4 (18:18 UTC) that Conductor is joining Meta, where the team will build "the next interface" with Nat Friedman and Alexandr Wang at Meta Superintelligence Labs (MSL). The official announcement sets a timeline: from November 4 there are no new sign-ups and Pro features are free for existing users; on November 13 Conductor Cloud goes read-only and active subscribers get refunds; on November 20 Cloud shuts down and cloud data deletion begins; the local app keeps getting new models and critical fixes until early 2027. Terms were not disclosed.

Conductor Is Joining Meta: The Multi-Agent Coding Tool Team Moves to MSL, Conductor Cloud Shuts Down on November 20

Key Takeaways

  • Conductor launched in July 2025 as a tool for running a team of coding agents at once (it runs first-party agents such as Claude Code, Codex, Cursor and OpenCode under the hood), and its site claims 100k+ builders.
  • The team joins Meta Superintelligence Labs to work with Nat Friedman and Alexandr Wang; the official word on the new product is only "early next year", with no details.
  • Key dates: sign-ups and downloads stop on Nov 4, Cloud goes read-only with refunds on Nov 13, Cloud shuts down and data deletion starts on Nov 20, and the local app is maintained until early 2027.
  • To migrate, right-click a cloud workspace and choose Transfer to local, or export everything via Settings, Account, Export data.
  • Terms were not disclosed; on the same day HumanLayer announced it is joining Factory, a sign the coding-agent tooling market is consolidating quickly.
Read full story
Linked Models & Agents:
🤖 Conductor🤖 Claude Code🤖 Codex🌐 Meta🌐 Conductor
#Conductor#Meta#MSL#Acquisition#coding agents#Multi agent
A
Axiom@axiommathai·
👀11
🔥84
View on 𝕏 ↗

Axiom Formalizes the Entire Stacks Project in Lean With AI Agent Swarms: About 16 Million Lines, Billed as the Largest Reusable Formal Math Library Yet

Axiom released a Lean formalization of the entire Stacks Project on November 4 (official X post at 17:13 UTC, with over 2 million views). The roughly 7,500-page open-source algebraic geometry reference, started by Johan de Jong in 2005, becomes about 16 million lines of code, which Axiom calls the largest integrated, reusable formal math library ever built. Hierarchically coordinated AI agent swarms did the work and surfaced dozens of statements in the original that needed local fixes, with counterexamples. The library is open source and browsable side by side with the original text at stacks.axiommath.ai.

Axiom Formalizes the Entire Stacks Project in Lean With AI Agent Swarms: About 16 Million Lines, Billed as the Largest Reusable Formal Math Library Yet

Key Takeaways

  • Scale: about 7,500 pages of source text and roughly 16 million lines of Lean, which Axiom calls the largest integrated reusable formal math library (vendor claim).
  • Method: hierarchical AI agent swarms translate statements, search Mathlib for existing abstractions and build proofs in parallel, with multi-agent auditing and semantic cross-checks against the source to keep statements faithful.
  • Findings: the system surfaced dozens of statements needing local refinement, including missing hypotheses and edge cases in positive characteristic, and built explicit counterexamples where claims failed.
  • The permanent Stacks tag structure is preserved so every Lean declaration links back to the original text; stacks.axiommath.ai offers side-by-side reading and a dependency graph.
  • This follows Axiom's October 27 formalization of the Classification of Finite Simple Groups (CFSG); per de Jong's request, corrections are kept in Axiom's own repository rather than sent upstream.
Read full story
Linked Models & Agents:
🌐 Axiom
#Axiom#Lean#Stacks Project#formal verification#AI agents#Math