Woven by Toyota: AI agents that fix MISRA code violations

Software for automated driving and driver assistance (ADAS) is written largely in C and C++ and must follow the MISRA coding guidelines, which run to hundreds of pages. Static analysis tools flag large numbers of violations, engineers fix them by hand, and fewer engineers now have deep C and C++ skills. One ADAS recognition module alone needed about 60,000 errors corrected.

Woven adds that the amount of code in a vehicle keeps growing, which makes manual compliance work harder to sustain.¹²

The tool, called MISRA Copilot, was built with partner Zen Architects. Three agents work in sequence: a Coder agent proposes compliant fixes, a Reviewer agent critiques them and sends them back for revision, and an Evaluator agent checks compliance and gives a rationale and a confidence score for each change. Woven says the agents draw on a knowledge base of earlier code fixes from Woven and Toyota through retrieval-augmented generation.

The tool never edits code on its own. It submits annotated proposals as GitHub pull requests, and engineers decide whether to accept them. Microsoft's story says the system runs on Azure OpenAI Service (now Azure OpenAI in Microsoft Foundry Models) with GPT-4o in early tests, o1 tested in January 2025 and o3-mini for the Coder agent, orchestrated with the open-source AutoGen framework, and uses Azure App Service, Azure Cosmos DB and GitHub Enterprise.

The team refined the system in two Azure hackathons in December 2024 and January 2025, and built a VS Code extension after a presentation to Toyota Group subcontractors at the end of January 2025.¹²³

In our view, the design travels well to any regulated codebase: one agent proposes a fix, a second critiques it, a third explains it and scores its confidence, and nothing reaches the code until an engineer approves a pull request. That keeps accountability with people while the repetitive work is automated. Readers should note what the numbers are: proof-of-concept results on a specific set of internal code. In November 2025 Woven said the tool was still in further evaluation for possible deployment, and the 80% in Microsoft's headline is a test figure, not a production rate. The cost projection has no published method. A team with modest violation counts, or one whose static analysis tool already offers reliable automatic fixes, may not need a multi-agent system.