Meta has built an internal AI agent that functions as a stand-in specialist for a compliance domain, and the company says it now handles the bulk of routine expert assessments that used to take days. Meta engineers Shaurya Sengar, Jason Nawrocki, Jay Shah, and Prashant Kommireddi described the system in a company engineering blog post. The design splits into two layers that rarely appear together in production agent systems: a structured knowledge base kept separate from the reasoning steps that consult it, and an automated pipeline that edits that knowledge base whenever an expert corrects the agent, without touching the underlying model’s weights.
The knowledge layer holds more than 200 files organized into four categories: position files that state the organization’s stance on a given question, taxonomy files that fix vocabulary, routing indexes that decide which files apply to a given query, and gateway files that block the agent from applying specialized knowledge where it should not. Every file lists its dependencies and consumers in its metadata, so a single edit can be traced through the files it touches before it ships.
Separating knowledge from reasoning is the same design argument the agent-memory crowd has been making for smaller, single-user tools. Meta’s version deploys it at enterprise scale, where the payoff is not a tidier prompt but auditability: a change to a position file never has to touch the step-by-step procedures, called recipes, that decide when to load it, and a bad output can be traced to one layer or the other rather than diagnosed as a black box.
Recipes are the second half of the architecture. Where a knowledge file states a fact, a recipe prescribes a workflow: what to check first, which files to load, what a complete analysis looks like. Meta says restructuring from a flat instruction file into staged recipes cut the tokens loaded per query by roughly 80 percent, since each step now pulls only the material relevant to that phase instead of retrieving broadly on every turn.
The self-improvement loop is the part Meta’s team calls the most distinctive piece of the system. When a domain expert corrects the agent, the pipeline runs that correction through four stages: diagnosing the root cause, drafting a minimal edit, validating the fix against a replay of the original scenario plus a regression suite, and routing the approved change to a human for final review. A separate agent, given only the proposed edit and no context on why it was made, checks the diff for contradictions before it lands. Each fix that clears review then joins the regression tests, which stops a later edit from quietly reversing it.
Meta’s reported numbers cover six weeks of work across three development sprints: experts judged the agent’s output useful nearly every time, a single assessment shrank from days of work to minutes, and no improvement cycle introduced a regression. These are Meta’s own internal figures, measured against its own compliance benchmarks, not an independently audited comparison, and Meta has not disclosed the size of the expert team or the query volume behind those numbers.
The company frames the architecture as usable well beyond this one compliance domain, wherever specialist knowledge sits in people’s heads rather than in text: financial risk, security review, procurement evaluation. For any team already running a retrieval-based support or compliance agent, the narrower design argument worth testing is this: separate the facts an agent draws on from the steps it follows to use them, so a bad answer can be traced to one file instead of retrained away.
Meta detailed the system in a September 2, 2026 post on its engineering blog.
