AI agents migrated 40,000 lines of Fortran, with humans at every gate

Every organisation with a core system older than most of its staff asks the same question about AI: can it finally move the thing? Mistral has published a detailed answer, and its failures are more useful than its success.[1]

The French lab says it helped a European energy operator migrate 40,000 lines of Fortran 77 to C++. The code is a reservoir simulator, with no test suite and no central documentation. The first sprint covered 40,000 of 300,000 lines.[1]

Translating syntax between languages is largely solved, Mistral says. Migrating a full procedural system into object-oriented C++ is not, because it forces architectural refactors. Fortran 77 was standardised in 1977 and has no modules, no namespaces and no structured types.[1]

State lives in COMMON blocks, global memory shared across the entire program. Variables are implicitly typed by their first letter, so a misspelt name silently creates a new variable instead of raising a compiler error. Names are capped at six characters.[1]

Before any agent touched the code, the team built a parity harness. Agreement meant numerical equality of outputs: the final results and critical intermediate points flagged by the client's own reservoir engineers.[1]

They added subroutines to export the Fortran program's state, a C++ test framework to load those checkpoints, and instruction files to steer agents into using them. Mistral calls this a net-positive investment that should come first in any modernisation job.[1]

The documentation sat in old PDFs and in comments buried inside the Fortran itself. The team parsed the codebase into a caller-callee tree, then used Mistral's Vibe command line tool to spawn more than a hundred agents to document it.[1][2]

Work started at the leaves of the tree and moved upward. Each node spawned a subagent that documented it and opened a pull request. A reviewer agent ran on a schedule, checking those requests and scheduling fixes. Mistral calls that documentation pass one of the largest side wins.[1]

The first attempt gave one agent per Fortran subroutine a week of full autonomy. The result was functional, but COMMON blocks became global structs one for one and GOTO control flow stayed intact. Mistral calls it Fortran retyped in C++ syntax.[1]

The second attempt added structure: a planner, a coder, a tester and a code quality reviewer on each module. Quality improved, then the agents hit bugs, tried a few fixes and stalled with nobody to intervene. The version that shipped puts a human in that gap.[1]