In a mid-October Engineering at Meta post, Meta described how a small team built Forum, a standalone app for deeper conversations in Facebook groups, by treating an AI agent as a participant in the development cycle rather than a faster typist. The official Engineering at Meta account later pointed to the post on X.
Success, Meta says, wasn’t measured in code volume, but in “verified, production-ready outcomes” — changes that build, pass tests, and survive review.
“It’s also where we tested a different way of working: treating an AI agent as a participant in the engineering process rather than a faster way to type.”
— Engineering at Meta, in How Meta Built Forum With AI-Native Engineering
The post contrasts human-in-the-loop, where the engineer drives every step and the AI assists inside it, with human-on-the-loop, where the agent runs the iteration — plan, implement, build, test, correct, repeat — and the engineer sets the goal, watches the run, steps in when it drifts, and approves the result. Meta calls human-on-the-loop “the target we are driving toward, not a finished state,” and says every shipped outcome still carries an engineer’s sign-off.
Meta says 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. An autonomous stress-testing loop, it says, caught a memory issue during development, roughly halving memory use and page faults.
To keep agents from getting lost in the interface, the team gave them tools to read the live view hierarchy from a simulator, rather than descriptions of the screens. As the app neared production, a separate sandbox build compiled prototype code out of the production build entirely, so experimental work couldn’t ship by sitting inside a feature flag.
After launch, agents help investigate user feedback and forward the analysis to engineers, who decide what to do next.
Meta also listed what didn’t work as well. Context got lost across sessions unless architectural decisions were written down. Agents still produced pull requests that technically passed while cutting corners, so quality checks had to be strong enough to catch the shortcuts. And remote iOS simulator connections were unreliable enough that local development was faster.
The same engineering blog has also described JiTTesting, a related human-on-the-loop approach to catching bugs with failing tests.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
