Google DeepMind wants AI Agents to produce code alongside formal proofs of correctness.
Lean can help verify whether software meets precise mathematical specifications.
Formal Verification adds stronger assurance than testing alone, but depends on accurate specifications.
Google DeepMind is pushing toward a new model for software development: AI systems that do not just write code, but also produce formal proof that the code meets clear safety, security, and correctness rules.
A current DeepMind research role for ‘Verified Code Generation’ calls for AI systems that can create formal proofs with the Lean proof assistant. The work also covers formal program specifications, verified static analysis, and code generation for large real-world software projects.
The idea addresses a major weakness in AI coding tools. A language model can produce code that looks correct and can even pass many tests. Yet hidden errors can remain. Formal verification takes a stricter route. It describes what a program must do in precise mathematical terms, then uses a proof system to check whether the program meets those rules.
Formal verification may sound complex, but the basic idea is simple. A developer first states a rule that the software must follow. The AI can then create code and a mathematical proof that connects the code to that rule. A trusted proof checker examines the result.
Lean plays a key role in this approach. Lean is a proof assistant that can represent mathematical statements and check every step of a proof. Google DeepMind has already used Lean for mathematical reasoning. Its AlphaProof system was trained to prove mathematical statements in Lean, which gave DeepMind an early foundation for AI systems that work with machine-checkable reasoning.
The newer verified-code effort moves that idea from mathematics toward software. Instead of asking an AI tool only to produce a function, the target becomes a pair: working code and a proof that the code satisfies its formal specification.
AI has made software creation much faster, but speed creates a difficult question about trust. If an AI agent writes thousands of lines of code, human experts cannot examine every detail with the same depth as a small manually written program.
Formal verification offers another layer of protection. Tests check selected cases. A formal proof can establish a stated property across all cases covered by the specification. That difference matters most for software that controls important systems, handles sensitive data, or supports AI infrastructure.
Google has also worked on CodeMender, an AI agent that finds and fixes software vulnerabilities. The project shows a wider DeepMind interest in AI tools that can improve software security rather than only generate new code.
Lean has gained a larger role in AI research as proof systems become part of the development process. DeepMind’s recent work also highlights a broader idea: an AI agent can create a proof, express it in Lean, and let a computer check whether the proof holds. That gives a clear answer about the validity of a formal claim.
The same model can apply to software. An AI system may suggest an implementation, generate the required proof, receive an error from the proof checker, revise the code or proof, and try again. Such a loop can turn verification from a separate expert task into part of the AI development process.
Also Read - Master Large Language Models in 2026: 10 Must-Visit GitHub Repositories
Formal verification does not solve every AI safety problem. One major difficulty comes before the proof itself: the system needs a correct specification.
A proof can show that code follows a rule. It cannot decide whether that rule captures every requirement that matters. If a specification misses an important security condition, a perfect proof cannot repair that gap.
Recent research highlights this problem. A 2026 study found that specification generation remains a major challenge for large language models. Precise specifications require more than simple translation from human language into formal language.
This issue matters for AI agents in particular. Their behavior can involve tools, external systems, changing data, and long sequences of actions. A short formal rule may not capture every risk.
The field has now moved beyond small examples. A new benchmark called Vero, published in August 2026, tests whether AI agents can create both software and formal proofs across complete multi-module repositories. The benchmark contains 43 instances from real-world repositories and covers areas such as cryptographic protocols and distributed systems.
The results show how early the technology remains. The strongest tested agent solved only 27 of the 43 instances. It also failed to close the specifications on the hardest repositories.
That result puts Google’s verified-code work in context. The goal is ambitious, but repository-scale verification still has major technical barriers.
Google’s work fits a larger change across AI research. Developers once focused mainly on better model output. The next stage adds systems that can check that output through external rules, tests, formal proofs, or other reliable methods.
Formal verification offers one of the strongest forms of this approach when the target property can receive a precise mathematical definition. A language model may make mistakes, yet a proof checker rejects an invalid proof regardless of how convincing it looks.
Google’s Secure AI Framework also takes a broader security approach across the AI development lifecycle, with controls for model risk, security, and privacy. Formal verification does not replace that framework. Instead, it can serve as a deeper technical method for specific software properties.
Also Read - Google to Roll Out Play Store Age Verification Globally by 2026
The long-term goal is clear: AI should help create software that comes with machine-checkable evidence of correctness, safety, and security. Google DeepMind’s current verified-code research shows that this idea has moved into active engineering and research work.
The technology still faces serious limits. Formal specifications remain difficult to create. Complex repositories remain hard for AI agents. Real-world AI safety also covers many issues that software proofs cannot capture.
Still, the direction marks an important change. AI code generation may no longer stop at 'the code looks right' or 'the tests passed.' The next standard could ask for something stronger: the code works, the required property has a precise definition, and a trusted system can check the proof.
1. What is Google’s Formal Verification approach?
It aims to combine AI-generated code with formal specifications and machine-checkable proofs.
2. What role does Google DeepMind play?
Google DeepMind researches AI systems that can generate verified code and formal proofs.
3. What is Lean?
Lean is a proof assistant that checks mathematical statements and proofs with high precision.
4. Can Formal Verification make AI completely safe?
No. It can verify defined properties, but it cannot guarantee that every important risk has a suitable specification.
5. How could this help AI Agents?
AI Agents could use verified code and proof checks to reduce certain software errors and security risks.