How Meta Built Forum With AI-Native Engineering
Engineering at Meta
Meta built Forum with an AI agent in the loop, not just as a code helper. The bet: verify everything before review, so the machine ships less fluff and more working code.
Based on reporting by Engineering at Meta — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta’s new Forum app is meant to pull Facebook group conversations into something people can actually go deeper on, and the way it was built is the real story. A small team started from scratch. They didn’t just use AI to type faster; they put an AI agent into the engineering process itself and judged it on verified, production-ready outcomes, not raw code volume.
That shift matters because Meta is trying to move from human-in-the-loop to human-on-the-loop. In the old model, the engineer drives every step and the AI fills in gaps. In the new one, the agent can plan, implement, build, test, correct, and repeat while the engineer sets the goal, watches the run, and signs off on the result. Not every part of that loop is fully autonomous yet, but the direction is clear: the person owns the outcome, while the agent does more of the mechanical work.
Meta’s answer to the obvious risk was verification, and lots of it. A task wasn’t considered done when it compiled. It had to pass automated tests and visual checks, with build results, test outcomes, and recorded verification sessions attached for review. The company also built navigation agents that act like users inside the live app, along with stress tests that mimic aggressive scrolling, random browsing, rapid navigation, and repeated media loading. Those runs helped catch a memory issue during development, and the fix roughly halved memory use and page faults.
The environment mattered as much as the prompts. Meta says the biggest gains came from making the codebase and tools easier for agents to reason about. When agents got lost editing the wrong screen or guessing coordinates, the team gave them direct tools to inspect the live view hierarchy and on-screen elements instead of relying on descriptions. Where prototype work needed freedom, Meta split out a sandbox build flavor that keeps experimental code out of production entirely while sharing the same codebase and design system.
The result, according to Meta, was a core feature-parity build in eight weeks. After launch, agents kept helping with root-cause analysis by turning user feedback into investigations for engineers to review. That’s the real pattern here: not AI replacing the team, but a small team using AI to compress the boring parts and keep the quality bar high enough to survive contact with real users.
My take — AI-written commentary, not fact-checked reporting
This is the least silly AI coding story in a while, because Meta treated the agent like a junior that needs supervision, tools, and guardrails. That is also why it worked. The industry keeps selling magic typing, when the real prize is less rework and fewer fake wins from code that merely compiles.
Read more about this at: Engineering at Meta