Meta ships Forum in eight weeks with agent-run engineering loop

Meta frames Forum as an experiment in how engineering work is organized, not just a new app. Its distinction is between human-in-the-loop, where an engineer drives every step and AI autocompletes inside it, and human-on-the-loop, where the agent runs the cycle and the engineer owns the outcome. Meta is explicit that it has not arrived: some stages run autonomously today, others still keep an engineer in each iteration.

The account also describes a trade-off in how code gets produced. Meta says it set the bar at verification rather than compilation, so a task counts as complete only after automated tests and visual checks pass. In its telling, agents check their own work before an engineer sees it, which means less is generated and less is discarded, and the engineer's sign-off carries attached build results, test outcomes, and recorded verification sessions. Meta says the limit on how much can be handed to the loop is how much the verifiers can be trusted.

Much of the piece concerns adapting the environment rather than the prompts. Meta rebuilt some architecture from scratch, trading reuse for agentic throughput, and gave agents tools to read the live view hierarchy and list on-screen elements instead of describing the app to them, which it says sharply cut misdirected edits. Forum's own feed work kept a human decision: engineers chose among the architectural options agents mapped, because deciding what a good outcome looks like is a product judgment.