How Meta Built Forum With AI Agents Across the Whole Development Cycle

An Agent as a Participant in the Build

Meta's newest app, Forum, is a standalone Groups app built by a small team from scratch. It is designed for people who want to go deeper into the conversations in their Facebook groups. The project also served as an experiment in treating an AI agent as a participant in the engineering process rather than a faster way to type. The team measured success not by code volume but by verified, production-ready outcomes: changes that build, pass tests, and survive review.

The shift is described as a move from human-in-the-loop to human-on-the-loop. In the older model the engineer drives every step and the AI assists inside it. In the newer model the agent runs the iteration itself, planning, implementing, building, testing, correcting and repeating, while the engineer sets the goal, watches the run, steps in when it drifts and approves the result. The team calls this a target rather than a finished state: some stages run autonomously today, while others still keep an engineer in each iteration.

Verification Became the Finish Line

The engineers set the bar at verification rather than compilation. A task is complete only once automated tests and visual checks have passed, and agents check their own work before an engineer sees it. Shipped changes carry build results, test outcomes and recorded verification sessions alongside an engineer's sign-off.

Four capabilities let the loop run safely on its own: planning treated as technical specification, automated verification signals such as build status and UI spec matching inside the loop, simulated user testing with navigation agents that interact with the live app and report regressions, and agentic production operations where agents analyze user feedback for possible regressions and hand the analysis to engineers.

Rules, Tools and a Safe Sandbox

The operating model rests on five components. Rules are permanent project instructions, encoded once when an agent repeats a mistake more than twice. Skills turn team conventions into loadable steps. Agents automate checks that previously needed a manual pass. Commands chain skills and agents into workflows that can assess complexity, reproduce a bug, plan and implement a fix, open a pull request and run a multi-stage review. Hooks are automated triggers, such as running a linter after every agent turn.

The biggest gains came from changing code and tooling rather than prompts. The team kept the codebase modular and legible, rebuilding some parts from scratch where reuse blocked agentic throughput. Instead of describing the app to the agent, they gave it tools to read the live view hierarchy, list on-screen elements and look up the path to a target view, which sharply reduced misdirected edits. As Forum matured toward production, a separate sandbox build flavor let prototypes run on the same codebase while keeping prototype code physically out of the production build.

Results: Parity in Eight Weeks, Cleaner Code

Forum reached core feature parity with Facebook Groups in eight weeks, with AI handling much of the routine implementation so engineers could focus on architecture, product decisions and quality. Automated stress scenarios that mimic real user behavior caught a memory issue during development; the agent isolated a repro, proposed a fix and verified it, roughly halving memory use and page faults. After launch, investigation agents triaged user feedback and posted structured write-ups for engineers to validate. The team credits context, not better prompting, for the quality gains, noting that training an agent once means every future session starts from it.

Sources: Engineering at Meta