How Meta Built Forum With AI-Native Engineering
Shengbo Guo, Bohou Luo, Shubhojeet Sarkar, Vivian Qu, Max (Xiangjun) Fan, Prerna Jain, Mike Plumpe, Chinmay Abhyankar, Rish Tandon · 15 October 2026
Read the originalMeta's Forum team let AI agents plan, build and test changes while engineers set the goal and signed off, and matched Facebook Groups' core features in eight weeks.
Forum is a standalone app for Facebook groups that a small team at Meta built from scratch. Its engineers moved from steering every step of an AI coding agent to setting the goal, watching the run and approving the result. A task counted as done only when automated tests and visual checks had passed, not when the code compiled.
- Every shipped change still carried an engineer's sign-off, with build results, test outcomes and recorded verification sessions attached to the pull request.
- When the agent repeated a mistake more than twice, the team wrote the correction once into permanent project rules for everyone.
- Agents kept editing the wrong screen until the team gave them tools to read the live app's view hierarchy and list what was on screen before acting.
- A separate sandbox build let designers and partners prototype with AI tools without an engineer for each change, and kept prototype code out of the production app entirely.
An engineering team handing more of each change to agents can make passing tests and visual checks the definition of done, so that reviewers see fewer, better-evidenced changes. This is Meta's own engineering blog describing its team's work.
Count an AI agent's task as done only when automated tests and visual checks pass, and attach the results to the pull request for the engineer who approves it.
When an AI agent repeats the same mistake more than twice, write the correction into the project's permanent rules for the agent.
Derived by Working Surface from the article; more in the Playbook. Source line: Compiling isn’t the finish line — a task is complete only once automated tests and visual checks have passed.
21 October 2026: as AI tools multiply code, documents and prototypes, the work that counts moves to judging, reviewing and keeping track of what they produce.