Modernizing complex legacy code with AI agents.

Mistral helped a European energy operator migrate 40,000 lines of Fortran 77 to C++ for a physics‑intensive reservoir simulator that lacked a test suite and centralized documentation. The migration required more than syntax translation; it demanded architectural refactors to move from procedural Fortran to object‑oriented C++. To ensure correctness, the team first built a parity harness that exported state snapshots from the Fortran code and used a C++ test framework to verify numerical agreement on critical outputs. This approach provided a cheap, convincing proof that modules were correctly migrated. Documentation, scattered across PDFs and comments, was reconciled and moved next to the code. A custom parser generated a caller‑callee tree, which agents used to document each subroutine. Initial attempts gave agents full autonomy, producing functional but unmodernized code. Subsequent iterations introduced structured roles—planner, coder, tester, reviewer—improving quality but still stalling on complex bugs. The final workflow combined human oversight with coder, tester, and reviewer agents, allowing module‑by‑module migration with human checkpoints. The first sprint covered core functionality, covering 40,000 of 300,000 lines. The project underscored three lessons: build a parity harness first, document before relying on agents, and use structured workflows with human gates for large legacy migrations. Mistral’s Applied AI team delivered the solution, illustrating how AI can support high‑stakes domain‑specific transformations.