Meta Figured Out That Better Context Beats Better Models, and Published the Receipts

Meta's engineering blog published something this week that I think every team shipping with coding agents needs to read. They had a problem: AI agents couldn't navigate their large-scale data pipeline codebases. The popular answer would be "use a smarter model." Meta went the other direction entirely.

They deployed 50+ specialized AI agents that systematically read every file across 4,100+ files in three repositories, producing 59 concise context files that encode tribal knowledge. Not code comments. Not documentation. Tribal knowledge. The stuff that lives in senior engineers' heads: why this module exists, what breaks if you change it, which config values are load-bearing.

The results are striking. Coverage jumped from 5% to 100% of modules. Preliminary tests show 40% fewer tool calls and tokens per task. That's not a marginal improvement. That's a fundamentally different cost profile for running agents at scale.

What I find most interesting is that this approach is model-agnostic. The context files work with any model you throw at them. It's the same principle behind CLAUDE.md files, AGENTS.md, cursor rules, or any project-level context injection. The insight is that the bottleneck for coding agents isn't reasoning capability. It's knowing where to look.

I've been doing a version of this in my own projects. Every repo I work in has a CLAUDE.md with architecture decisions, file locations, and conventions. It's not glamorous work. But it's the difference between an agent that flails for 20 tool calls trying to understand the codebase and one that gets to work immediately.

Meta also dropped KernelEvolve in the same period. It treats kernel optimization as a search problem, exploring hundreds of Triton kernel implementations to find solutions matching human expert performance. 60% throughput improvement in hours versus weeks for humans. Accepted at ISCA 2026. The common thread: Meta is treating AI agent effectiveness as a systems problem, not a model problem.

If you're running agents against any codebase larger than a toy project, build your context files this week. Start with a single markdown file per module documenting what it does, why it exists, and what depends on it. Your agents will immediately get faster and cheaper.