Meta’s newest app, Forum — a standalone experience for people who want to go deeper into conversations in their Facebook groups — was built by a small team that put an AI agent into the development cycle as a participant rather than a faster typist, the company wrote on its engineering blog on 15 October 2026.
Forum is described as built from scratch. Meta says the experiment needed a clear success definition that was not code volume: “verified, production-ready outcomes — changes that build, pass tests, and survive review.”
Human-in-the-loop versus human-on-the-loop
Meta’s post contrasts two modes.
With human-in-the-loop, the engineer drives every step and the AI assists inside it — autocomplete, a suggested fix, an answer to a question. Progress stays paced by the person.
With human-on-the-loop, the agent runs the iteration itself — plan, implement, build, test, correct, repeat — while the engineer sets the goal, watches the run, intervenes when it drifts, and approves the result. Meta writes: “This is the target we are driving toward, not a finished state: some stages run autonomously today, while others still keep an engineer in each iteration.”
Every shipped outcome, Meta says, still carries an engineer’s sign-off. The investment was in making that sign-off fast and well-evidenced, with build results, test outcomes and recorded verification sessions attached to the change.
Verification inside the agent loop
Meta argues that compiling is not the finish line: a task is complete only once automated tests and visual checks have passed. The team integrated automated pass/fail signals — build status, unit tests and UI spec matching — directly into the agent loop so agents check their own work before an engineer sees it.
Four capabilities, in Meta’s account, make unsupervised cycles safer:
- Planning as technical specification — rigorous upstream decomposition and machine-readable context, because ambiguity that a human might resolve by intuition causes agent failure.
- Automated verification in the loop — the force multiplier for ready-to-review pull requests with pre-populated test plans.
- Simulated user testing — navigation agents that use the live app “exactly like a person would,” performing actions, reading the screen and reporting regressions.
- Agentic production operations — after launch, agents help with root-cause analysis from user feedback, then forward analysis to engineers who decide next steps.
Operating model: rules, skills, agents, commands, hooks
Meta describes machinery with five components: permanent rules (encode a correction once when the agent repeats a mistake); skills that package team conventions; agents that automate checks; commands that chain skills and agents into workflows (assess complexity, reproduce a bug, plan and implement a fix, open a PR, multi-stage review); and hooks such as linting after every agent turn or validating tests before a PR is created.
Adapting the codebase for agents
The biggest gains, Meta says, came from changing code and tooling, not prompts. The team modularised architecture so agents could plug in more easily, and sometimes rebuilt from scratch when reuse hurt agentic throughput. A recurring failure mode was the agent losing track of where it was in the UI; the fix was tools to read the live view hierarchy and on-screen elements rather than prose descriptions of the app.
For safe prototyping after Forum had real users, Meta used a separate build flavor — a sandbox app sharing the same codebase and design system as production, but compiling prototype code out of the production build via a plugin/socket architecture — so experiments could not crash the production binary.
What this post does not claim
Meta’s blog includes outcome anecdotes and velocity colour. Per TSN’s brief for this item, this article does not add productivity figures — no invented speed-ups, and no recycling of numeric “how fast / how much” claims from the blog into the news account. The news is the engineering method Meta says it used on Forum, and the human-on-the-loop target it is driving toward.
Why it matters
Large labs talk constantly about coding agents; fewer publish a full operating model that ties agent loops to build/test/UI verification and post-launch investigation. Meta’s Forum write-up is a concrete case study of treating verification, not token output, as the bottleneck — while still insisting on engineer sign-off for anything that ships.
For other product teams, the portable ideas are less about a specific model and more about environment design: legible architecture, tools that expose live UI state, sandbox builds, and pass/fail signals inside the loop.
What’s still unknown
- Which model or internal agent stack Forum’s team used.
- How far human-on-the-loop has spread beyond Forum inside Meta.
- How Forum’s product roadmap relates to the main Facebook Groups surface.
What to watch
Watch for follow-on Meta engineering posts that generalise the Forum operating model, and for Forum product updates that show whether the same agentic discipline continues after launch.
What this means
- Forum was built with agents in the loop. Meta’s newest Groups-focused app is the case study.
- Human-on-the-loop is the target, not the finished state. Engineers still sign off every ship.
- No productivity figures here. Method and verification bar only, per TSN’s brief.
Sources
Related stories
- Meta Releases the Weights for Muse Spark Open, and Apple Shows It Running Across Two to Four Mac Studios
- Meta Launches Muse Wonder and Muse Wonder Full, Its “Most Capable Models to Date”, Built for AI Agents
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
