Artificial intelligence November 12, 2026

Axiom Math claims AI agents verified the Stacks Project in Lean

Axiom Math says its AI agents have formally verified the Stacks Project, translating roughly 7,654 pages of mathematics into about 16 million lines of Lean code. The scale deserves attention: Stacks is an interconnected reference spanning algebra, ge...

Axiom Math claims AI agents verified the Stacks Project in Lean

Axiom Math’s Stacks claim puts AI-generated proofs to a harder test

Axiom Math says its AI agents have formally verified the Stacks Project, translating roughly 7,654 pages of mathematics into about 16 million lines of Lean code. The scale deserves attention: Stacks is an interconnected reference spanning algebra, geometry, and number theory, assembled by mathematicians over two decades.

TechCrunch’s report, published November 5, 2026, and updated that evening, says Axiom announced the work on November 4. The company’s project site presents the formalization.

The immediate value is a larger base of machine-checkable mathematics, particularly for algebraic geometry. Axiom also wants to give AI agents verified components they can reuse when writing software and eventually doing science. That ambition needs separate evidence. Translating a major mathematical library is a substantial accomplishment on its own; demonstrating improvements in production software is another task.

For engineers, the project raises a practical question: how much trustworthy work can accumulate when generation is cheap and a formal checker controls what gets accepted?

Keeping the mathematics connected

Columbia mathematician Johan de Jong started the Stacks Project in 2005 as an open-source reference. Contributors have expanded it from foundational algebra into advanced results across several fields.

Automated translation has to account for the library’s dependencies. Individual proofs rely on shared definitions and earlier results. A theorem might check successfully in isolation while fitting poorly with the abstractions used elsewhere. Across thousands of pages, those choices become an integration problem.

Lean lets authors express definitions, theorem statements, and proofs in a formal language. Its proof-checking machinery verifies that a proof establishes the stated proposition under the declared assumptions. Errors that can hide in fluent mathematical prose become detectable.

The translation still has to preserve the original mathematics. The formal theorem must say what the human theorem says, and its dependencies must reflect the intended foundations.

That’s why Axiom’s 16-million-line figure is useful mainly as an indication of scale. It doesn’t establish fidelity, maintainability, or how convenient the library will be for other researchers.

Axiom technical staff member Ashvin Swaminathan describes algebraic geometry as a major gap in existing formalized mathematics. If the company’s work faithfully covers Stacks and proves reusable, researchers will have a substantial starting point for checking new results, with less effort spent rebuilding prerequisites.

A difficult theorem may depend on a long chain of supporting mathematics. Formalizing that chain once allows later work to reuse it. Terence Tao describes this compounding effect in his discussion of AI and mathematics: “The more formalized proofs you have, the more formalized proofs you can make.”

Parallel agents still need shared abstractions

Axiom says it used a fleet of agents working in parallel, with a loose hierarchy rather than central assignment of every task. Agents identified work they could undertake themselves.

In principle, that suits a large dependency graph. Independent branches can progress concurrently, while shared definitions and foundational results determine where coordination is necessary. But scheduling is only part of the work. Agents also need compatible representations and interfaces.

Developers will recognize the problem: two teams produce reasonable modules that are painful to combine because they model the same concept differently. In formal mathematics, mismatched definitions can require extra translation proofs or substantial rework.

Axiom CEO Carina Hong told TechCrunch the team initially wasn’t sure whether its approach could preserve Stacks’ structure. That’s a credible concern. Scaling proof generation adds integration work alongside the search for individual proofs.

The company also says it trained its own mathematical models. Early on, an in-house Lean expert helped teach the agents mathematical “taste.” Hong says the agents later developed that judgment themselves.

Taste is an imprecise term for a concrete library-design problem. A proof can pass while using awkward definitions, unnecessarily specialized lemmas, or constructions that make downstream work harder. The checker enforces logical validity; choosing abstractions that other users can work with still requires judgment.

Technical leads building agent systems face the same constraint. Automated acceptance criteria can reject many bad outputs, but they don’t necessarily select the best architecture among those that pass.

Axiom’s account credits model training, expert guidance, and task coordination. The report doesn’t provide enough detail to isolate how much each contributed.

Risks and limitations

TechCrunch reports Axiom’s announcement and explains the company’s approach, but the supplied report doesn’t include an independent audit of the entire formalization. It also lacks a detailed coverage breakdown, compute costs, completion time, and figures for the human intervention required.

Without those details, engineers can’t assess efficiency or reproducibility. Sixteen million lines of output say little about the cost per accepted theorem or the effort needed to maintain the library.

The guarantee also has a boundary. Formal verification checks a formal statement under specified assumptions. It doesn’t automatically certify that the statement faithfully captures the original human claim.

A translation can weaken a theorem, encode a definition incorrectly, or introduce an assumption that changes what has been established. The proof may still check. Statements, definitions, and assumptions therefore need review as part of evaluating the formalization.

The source flags concerns about automated checking, linking to research on failures in checks of AI-produced mathematics. That supports caution about treating “computer-checked” as a complete quality label. It doesn’t establish that Axiom’s work contains such failures.

For a project this large, reviewers need to know which assumptions the proofs use, whether any obligations remain unproved, and whether another team can reproduce the checks using the published artifacts.

Axiom’s software ambitions remain an extrapolation. Swaminathan says the company believes the skills developed in mathematics transfer to building software infrastructure. The Stacks announcement doesn’t demonstrate that transfer.

Formal verification already has practical uses. The source cites chip-design verification and research that found bugs in software calculating French family benefits. Those examples establish its value in specific settings. They don’t show that an agent trained to formalize algebraic geometry can reliably verify an arbitrary web service.

What developers can take from it

Separating a flexible generator from a strict acceptance mechanism gives agent systems a useful working model.

An agent can attempt constructions, make mistakes, and revise its output. A formal checker tests whether a particular claim has been established. Accepted results can then become dependencies for subsequent work.

Software teams already use a weaker version of this approach with compilers, type checkers, tests, and static analysis. Formal proofs can offer stronger guarantees for explicitly stated properties, though writing those properties and modeling the system take effort.

Consider a hypothetical authorization component. Proving that a modeled operation requires a permission could be valuable. The guarantee still depends on whether the model captures the deployed permission rules and whether the running implementation corresponds to it. Configuration, external services, and operational behavior may fall outside the proof’s scope.

Bounded components are therefore a sensible place to evaluate the approach: code with clear contracts, explicit assumptions, and expensive failure modes. Teams need to know exactly which property was checked before accepting a claim that the software is “verified.”

Axiom, which TechCrunch describes as a $1.6 billion startup, raised a $200 million Series A led by Menlo Ventures in March 2026. It has substantial funding behind its bet that verified libraries will become infrastructure for agents.

Independent researchers building new formal proofs on the Stacks translation would provide useful evidence of its quality, especially if they report where integration gets difficult. For software engineers, the corresponding test would be a reproducible demonstration of agents producing verified code against a meaningful specification, with costs and human effort disclosed.

The library’s practical value will depend on how readily others can build on it, and how much inspection and repair they need to do first.

Keep going from here

Useful next reads and implementation paths

If this topic connects to a real workflow, these links give you the service path, a proof point, and related articles worth reading next.

Relevant service
AI agents development

Design agentic workflows with tools, guardrails, approvals, and rollout controls.

Related proof
AI support triage automation

How AI-assisted routing cut manual support triage time by 47%.

Related article
Artisan's $25M Series A puts AI sales agents in a narrower light

Artisan just raised a $25 million Series A for its AI sales platform, built around an outbound sales agent called Ava. The company is still hiring people too. That tracks. The slogan got attention. The actual product is narrower. Artisan broke throug...

Related article
Walmart’s Agent Gateway uses MCP to support AI shopping agents

Walmart is giving approved AI agents a supported way to search its inventory and buy products through a new MCP server called Agent Gateway. It’s also introducing account-level controls that let shoppers approve individual purchases or authorize spen...

Related article
Elad Gil on AI markets with real winners and the categories still open

Elad Gil’s read on the AI market is blunt and mostly right. Some categories already have leaders with real staying power. Others still look busy, funded, and vaguely promising, but nobody has earned the right to call them won. That distinction matter...