Mistral converted 40,000 lines of Fortran to C++. A human-directed agent worked best

Mistral published a case study on modernizing a physics-intensive reservoir simulator used by a European energy company. The original system contained roughly 300,000 lines of Fortran 77 and lacked both continuous documentation and a complete test suite. The team converted about 40,000 lines to C++. Before doing so, however, it created a verification environment that compared the numerical results of the old and new versions at final and selected intermediate points. The sequence of steps is the main lesson. Mistral recommends creating equivalence checks before agents begin the migration. In the project, more than a hundred agents then helped analyze the program's call tree, retrieve information from old documents and create documentation for individual components. The authors write that the verification environment was an unambiguous benefit to the project because it allowed long agent runs to be checked in a simple, measurable way. The team tested several levels of autonomy. A fully autonomous agent could produce a working result, but tended to preserve unsuitable elements of the original architecture. A more complex setup with planner, programmer, tester and reviewer roles ran into blocked workflows and high coordination costs. The best-performing approach was one in which a human determined the next step and the agent executed a bounded part of the migration. For companies with legacy systems, this is an important signal. The biggest obstacle is often not the syntax of a historical language, but knowledge of what the software actually does and which deviations are unacceptable. AI can substantially accelerate documentation, migration of repetitive components and test creation. It should not, however, make architectural decisions without people who understand the operation. The migration is therefore best divided into small modules with clear inputs, outputs and checks of numerical or functional equivalence. Clients should also change how such projects are procured and managed. Instead of a general objective to 'rewrite the system with AI,' it is better to define measurable contractual stages: mapping dependencies, creating reference tests, documentation, converting a specific module and proving equivalence. Time savings and error counts can be quantified at each stage. This separates the agents' real contribution from an impressive demonstration and reduces the risk that the company replaces old technical debt with new technical debt. Another consequence concerns ownership of knowledge. When an agent creates documentation, tests and a dependency map during migration, the company gains not only new code but also an asset that often did not exist before. This secondary output can have high value in audits, onboarding new developers or further component replacement. Modernization should therefore also be assessed by how much operational knowledge was converted into a maintainable form. KEY TERMS Fortran 77: An older programming language that still survives in many scientific and industrial systems. Numerical equivalence: Verification that the new program computes the same results as the original system. Technical debt: Costs created by long-term use of a solution that is difficult to maintain or obsolete.

Mistral converted 40,000 lines of Fortran to C++. A human-directed agent worked best

The full article is available to users registered at Hard Skills