How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines

AI coding assistants underperform on large, real-world codebases because they lack the organizational context that experienced engineers carry in their heads, rather than due to weak models. Meta encountered this problem when deploying AI agents on a data processing pipeline spanning four repositories, three languages, and 4,100+ files. The agents weren't making useful edits fast enough to justify the investment.

The root cause was tribal knowledge. Senior engineers know which modules own which logic, which abstractions are load-bearing, and which files are safe to touch. AI agents do not. They treat every file as equally relevant, burn context budget on noise, and miss the structural dependencies that only exist in team memory. Meta's fix was to encode that knowledge explicitly — building a codebase map that captures pipeline ownership, cross-repo dependencies, and domain-specific conventions that never appear in documentation.

With the map injected into agent context, the useful edit rate improved meaningfully. The specific numbers are not fully disclosed in the post, but the qualitative finding is unambiguous: the agents' limiting factor was contextual orientation rather than reasoning capability. The mechanism is straightforward: agents with a structured map spend less context budget on file exploration and more on the actual edit task. For a 4,100-file codebase across three languages, unguided retrieval is expensive and often wrong.

The honest limitation is that Meta's infrastructure team built this map semi-manually, meaning the approach does not automatically generalize. Constructing a high-quality tribal knowledge artifact requires engineers who already possess that knowledge to codify it, which represents a real upfront cost. Whether that cost amortizes well depends on how frequently agents interact with the codebase. For high-churn, high-volume development pipelines, it almost certainly does. For smaller codebases with lower agent usage, the ROI is less clear.

For teams encountering agents that generate syntactically plausible but contextually wrong edits, this points to a specific diagnosis. The problem is usually the absence of an explicit representation of how your codebase is actually organized and owned, not the model or the prompt.

Key takeaways:

• The bottleneck is tribal knowledge, not model capability. Agents exploring a large multi-repo codebase without orientation maps spend context budget on the wrong files and miss cross-repo dependency structure entirely.
• This implies current AI coding agents are highly sensitive to how well organizational context is externalized. Codebases with strong documentation and explicit ownership metadata will see disproportionately better agent performance.
• Teams deploying AI coding agents on large monorepos or multi-repo systems should invest in building explicit codebase maps before tuning prompts or swapping models, as the orientation artifact may deliver more lift than either.

Source: How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines