How Meta Used AI to Map the Tribal Knowledge of a Massive Data Pipeline

AI coding agents can't quickly make useful changes if they don't understand the codebase. To solve this problem in a massive data processing pipeline spanning 4 repositories, 3 languages, and 4,100+ files, Meta built a pre-compute engine that reads files first and structures the knowledge before agents act.

The core idea was to deploy 50+ specialized AI agents to systematically read every file and compress the tribal knowledge that previously existed only in engineers' heads into 59 context files. As a result, AI agents now have structured navigation guides for 100% of code modules, with coverage increasing from about 5% to 100%.

Each module was organized around 5 questions:

• What does it configure
• How do you modify it
• What breaks the build
• What modules does it depend on
• What undocumented knowledge is hidden in code comments

This process surfaced 50+ non-obvious patterns. For example, referencing intermediate-stage temporary field naming conventions incorrectly causes code generation to fail silently, and some deprecated enum values must not be removed due to serialization compatibility—rules like these were uncovered.

The generated context files followed a "compass, not encyclopedia" principle. Each file was kept to 25-35 lines, roughly 1,000 tokens, containing only 4 sections: Quick Commands / Key Files / Non-Obvious Patterns / See Also. All 59 files together take up less than 0.1% of a modern model's context.

The effect was also confirmed numerically. A self-refreshing system was attached that includes file path validation, coverage gap detection, critic re-runs, and automatic fixes for stale references, and in initial tests, tool calls and tokens per task decreased by about 40%. Also, across 6 tasks, the time needed to understand complex workflows dropped from about 2 days to 30 minutes.

However, Meta also notes research showing that AI-generated context files actually degraded performance on typical open-source codebases. That's why this approach was designed to be short, loaded only when needed, and quality-filtered through multi-round critic review. The key point is that this is especially effective not on public code the model already knows, but on proprietary config-as-code that the model doesn't know.