The problem has a name in every large engineering organization. Tribal knowledge.
It’s the understanding of a codebase that lives in engineers’ heads, the non-obvious design decisions, the workarounds, the reasons a particular pattern exists, that never makes it into documentation. When an AI coding assistant encounters a large, complex repository without that context, it makes edits that don’t hold. It calls tools repeatedly trying to orient itself. It produces suggestions that work in isolation and break in integration.
Meta built a system to fix this. The results, published today on Meta’s Engineering Blog by Krishna Ganeriwal, Plawan Rath, and Ashwini Verma, are worth paying attention to.
By the numbers
– 50+ specialized AI agents deployed to read the full codebase – 4,100+ files across four repositories and three programming languages – 59 concise context files produced, encoding tribal knowledge for each code area – 100% of code modules now have structured navigation guides (up from 5%) – ~40% reduction in AI agent tool calls per task, Meta’s own preliminary measurement*
*The 40% figure is self-reported from preliminary testing. It has not been independently validated.
How the system works
The core architecture is a pre-compute engine. Rather than waiting for an AI coding assistant to figure out a codebase in real time, the approach most current tools use, Meta’s system runs a swarm of agents ahead of time to read every file systematically.
Those agents produce 59 context files. Each file encodes what a human expert would know about that section of the codebase: navigation structure, non-obvious patterns, the design decisions that explain why the code looks the way it does. When an AI coding assistant is later invoked, it has those context files available as structured guidance rather than raw code it must interpret from scratch.
The codebase in question covers four repositories, three languages, and more than 4,100 files. Before this system, AI agents had navigation guides for 5% of those modules. Now they have guides for all of them.
Why developers and AI platform teams should care
This is a practitioner-grade case study from an engineering organization operating at scale. The tribal knowledge problem isn’t unique to Meta. It exists in any codebase large enough that no single engineer understands the whole thing.
The approach Meta describes is model-agnostic in design, the context files work as input to whatever AI coding assistant is in use. That portability matters. It means the architecture isn’t tied to a specific model release or API, which makes it more durable as the model landscape shifts.
The 40% reduction in tool calls has direct cost and performance implications. Every tool call is latency and, depending on the API, cost. An agent that needs fewer calls to complete a task is faster and cheaper to run.
What to watch
Meta hasn’t published this as a general-purpose tool, it’s a case study of an internal system applied to a specific codebase. The open question is whether the architecture generalizes. Does it work as well for codebases in different domains? What’s the cost of running 50+ agents for the initial pre-compute pass? How does the system handle rapid codebase evolution?
These questions aren’t answered in today’s post. Watch for follow-up publications from the named authors and for whether other frontier engineering organizations describe similar approaches.
TJS synthesis
Meta’s tribal knowledge system is one of the clearest published examples of a pre-compute context architecture for enterprise agentic AI. The approach trades upfront computation for runtime performance, an architectural bet that pays off when the codebase is stable enough to justify the investment. For AI platform teams building or buying agentic coding tools in 2026, this is the kind of first-party case study that should inform both build-vs-buy decisions and evaluation criteria. The metric to watch isn’t benchmark performance on generic tasks. It’s performance on your specific codebase. Meta’s system addresses exactly that gap.