Loading post…
MetaShengbo GuoProblem & solution
How Meta Built Forum With AI-Native Engineering
Summary
Meta built Forum using AI-native engineering, treating an AI agent as a development participant rather than a faster typing tool. The team shifted from human-in-the-loop to human-on-the-loop, where the agent autonomously plans, implements, builds, tests, and corrects its work while the engineer sets goals and approves outcomes. Success was measured by verified, production-ready results, not code volume, with automated verification signals like build status and unit tests filtering output before review. Meta adapted the codebase for legibility, gave agents tools to inspect live view hierarchy directly instead of descriptions, and created a separate sandbox build to isolate prototype code from production. This approach helped Forum reach core feature parity with Facebook Groups in eight weeks and caught a memory issue during development that roughly halved memory use and page faults.
Context
Forum is a standalone Groups app that a small team built from scratch. It served as a testbed for treating an AI agent as a participant in the engineering process, shifting from human-in-the-loop to human-on-the-loop where the agent drives iteration and the engineer oversees outcomes.
Approach / What changed
The team adopted human-on-the-loop: the AI agent autonomously executes the plan-implement-build-test-correct cycle while the engineer sets goals and approves results. To support this, they adapted the codebase for architectural legibility and modularity, gave agents direct state-inspection tools (e.g., live view hierarchy) instead of descriptions, and created a separate sandbox build flavor to isolate prototype code from production. They integrated automated verification signals (build status, unit tests, UI spec matching) into the agent loop, deployed navigation agents for simulated user testing, and after launch used agents for automated root-cause analysis from user feedback.
Takeaways
Success was measured by verified, production-ready outcomes, not code volume; AI-generated code had to pass automated tests and visual checks before reaching review.
Giving agents tools to directly inspect live app state (such as reading the view hierarchy) instead of using descriptions dramatically reduced misdirected edits.
Using a separate sandbox build flavor that compiles prototype code out of the production build allowed safe, fast AI iteration without risking the live app.
FAQ
How does human-on-the-loop differ from human-in-the-loop?
In human-in-the-loop, the engineer drives every step and AI assists within each step. In human-on-the-loop, the agent autonomously runs the iteration (plan, implement, build, test, correct) while the engineer sets the goal, monitors, and approves the outcome. The engineer shifts from executing steps to owning the quality bar.
How did Meta ensure quality of AI-generated code?
Meta set the bar at verification rather than compilation. Agents check their own work before an engineer sees it, requiring automated test and visual check passes. Only changes that build, pass tests, and survive review are shipped, with build results, test outcomes, and recorded verification sessions attached.
What was the sandbox build used for?
It was a separate build flavor (sandbox app) that shares the same codebase but compiles prototype code out of the production build entirely via a plugin/socket architecture. This allowed designers and engineers to prototype freely with AI tools without affecting the production app, and any sandbox result could be ported directly to production.
Source and freshness
- Source
- engineering.fb.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.