Meta uses AI to turn tacit knowledge into a code map

On April 6, 2026, Meta described an attempt to turn its engineers’ tacit knowledge into a context layer that its coding agents can consult. An assistant can find a function, propose a change and make the code compile; the hard part is knowing what it must not change. In large systems, much of that answer does not live in the repository: it is spread across conventions, old decisions and the memory of people who have maintained the product.

The case starts with one of its data-processing pipelines. According to the company, it spans four repositories, three languages and more than 4,100 files. A seemingly small change, such as onboarding a data field, can cross configuration registries, routing, graph composition, validation, code generation and automation. At first, the agents explored and guessed: they could produce changes that compiled but were wrong because of relationships the code did not make obvious.

A map before editing

Rather than asking one agent to summarize the repository, Meta prepared a pre-compute phase. More than 50 specialized tasks divided the work: explorers to map the code, module analysts, writers, independent critics, fixers, prompt tests and integration reviews. The result was 59 context files and a cross-repository dependency index.

Each file follows what Meta calls a “compass, not encyclopedia” principle. In roughly 25 to 35 lines, it gathers quick commands, the few key files, non-obvious patterns and cross-references. The aim is not to pour the system’s history into a model’s context, but to help it start its investigation in the right place.

That matters because the most valuable knowledge is often negative: values marked as deprecated that cannot be removed because of serialization compatibility; two modes using different names for the same operation; or temporary fields whose name changes later in a pipeline. Meta says it identified more than 50 such patterns. The system periodically rechecks paths, coverage gaps and stale references, because an outdated guide can be worse than no guide.

The figures come from an internal case

Meta says it moved from five files and about 5% coverage to 59 files covering every module in the pipeline. In preliminary tests across six tasks, agents with the pre-computed context used roughly 40% fewer tool calls and tokens. The company also says its critic passes verified referenced file paths without hallucinations.

Those results are useful for understanding the hypothesis, but they are not enough to turn it into an engineering rule. This is a proprietary system with specific knowledge that was not part of models’ public training, and the comparison covers six tasks. Meta does not present an external benchmark or an independent measure of the quality of deployed changes.

Context can also get in the way

A preprint by ETH Zürich researchers and collaborators provides a useful counterpoint. Evaluating repository context files on SWE-bench Lite and a custom task set, the authors found that model-generated files did not generally improve resolution rates and raised inference cost by between 20% and 23%, depending on the benchmark. Developer-written files performed somewhat better, but also increased steps and cost.

The difference helps explain Meta’s design. The point is not to add a long summary to every project, but to recover information that a model cannot know, load it only when relevant, and submit it to review. Even then, teams need to measure whether it reduces wasted exploration without imposing unnecessary rules.

The practical lesson is not that every repository needs dozens of agent documents. It is that documentation meant for a coding assistant must be specific, verifiable and kept current. In a complex system, a good map does not replace an engineer’s judgment; it reduces the time needed to discover where not to go.