What Happened

Meta's engineering team ran into a concrete problem: AI coding agents pointed at a large internal data pipeline — spanning 4 Python, C++, and Hack repositories with 4,100+ files — were too slow and too inaccurate to be useful. The agents lacked context about how the system actually worked.

Their fix was a pre-compute engine: a swarm of 50+ specialized AI agents that read every file and produced 59 structured context files encoding what Meta calls "tribal knowledge" — design decisions, non-obvious patterns, and cross-repo relationships that previously existed only in engineers' heads.

  • Navigation guides now cover 100% of code modules, up from 5%
  • 50+ non-obvious patterns documented (e.g., two config modes using different field names for the same operation)
  • Preliminary tests show 40% fewer AI agent tool calls per task
  • The knowledge layer is model-agnostic and works with most leading LLMs

The system is also self-maintaining: automated jobs run every few weeks to validate file paths, detect coverage gaps, re-run quality critics, and fix stale references.

Why It Matters

Most teams using AI coding assistants hit the same wall Meta did — the model is capable, but it has no map of your specific codebase. This article shows a reproducible architecture for solving that. The pre-compute approach separates knowledge extraction from task execution, which means you can swap models without rebuilding your context layer. For indie devs and SMEs, the key insight is that investing upfront in structured codebase documentation pays off in agent efficiency, not just human onboarding.

Asia-Pacific Angle

Chinese and Southeast Asian dev teams building on multi-repo microservice architectures — common in fintech, logistics, and SaaS platforms across the region — face the same tribal knowledge problem at scale. Teams using Qwen, DeepSeek, or other locally-deployed models can apply this exact pattern: build a pre-compute context layer once, then reuse it across model versions. This is especially relevant for teams going global who need to onboard new engineers or AI agents into complex legacy codebases without relying on senior engineers as a bottleneck.

Action Item This Week

Pick one module in your codebase where AI suggestions are consistently wrong or require heavy correction. Write a single context file documenting its non-obvious patterns, naming conventions, and cross-module dependencies. Test whether your AI assistant produces better edits with that file included in the prompt context. That's the manual version of what Meta automated.