Meta needed 50+ AI agents to map their tribal knowledge. The most durable piece of their stack is the part you can build today.

Short summary

Meta used 50+ AI agents to map tribal codebase knowledge, but the true innovation—buried in their post—is a deterministic dependency index that cuts common queries from 6,000 to 200 tokens. Unlike LLM-generated context files that decay and require continuous maintenance, this graph-based approach is self-maintaining and scales without a dedicated platform team.

• •Deterministic dependency graphs outperform LLM-generated context files for codebase queries
• •30x token reduction on common "what depends on X?" questions (6,000 to 200 tokens)
• •Graph-based systems self-maintain via code parsing; LLM artifacts require constant refreshing

Generated with AI, which can make mistakes.