How Meta built Forum with AI agents: the article and my margin notes (#AI4SDLC)

Meta, which is banned in Russia, has published an article about building Forum, a standalone app for Facebook Groups. An interesting case: a small team was building a new product and reshaping its work around AI agents at the same time. The article raises many familiar questions about adopting AI across the SDLC, which I marked in the margins.

What the authors describe

The team is moving toward human-on-the-loop development: the agent runs the “plan → code → build → test → fix” cycle, while the engineer sets the goal, intervenes when it drifts, and approves the result. This remains a target model: some stages still require a person in every iteration. An engineer signs off on every shipped change.

They built four foundations for this: precise task definitions, automated checks inside the loop, agents for testing user scenarios, and investigations of problems after launch. Changes should arrive at review with test results and recorded verification sessions. The team organized its working conventions into rules, skills, agents, commands, and hooks, ranging from permanent instructions to automatically triggered checks.

The team says changes to code and tools helped the most. They simplified the architecture, gave agents access to the simulator’s live UI state, and created a separate build for experiments. Stress tests exercised aggressive scrolling, navigation, and media loading; an agent could reproduce a discovered problem, investigate it, and propose a fix.

Meta reports that the team reached core feature parity with Facebook Groups in eight weeks, and that fixing one issue roughly halved memory use and page faults. These are results from a single project: the article provides neither a controlled comparison nor the full cost of preparing the environment.

My notes and points of focus

🔸 The whole SDLC and the definition of success The authors establish their measure of success at the outset: a verified change ready for production. This echoes a point I make when discussing AI adoption across the SDLC: what we do before coding and how we verify afterward both matter. Parallel agent work is useful as long as the engineer can trust what arrives for review.

🔸 Specifications and verification I wrote in the margins: “Their own version of spec-driven development?” There is decomposition and machine-readable context, although the authors do not describe a complete SDD methodology. More reliable checks inside the loop mean more work can be delegated to the agent. But the closing section contains an important qualification: agents managed to cut corners even in PRs that passed the checks. The definition of “done” itself therefore needs scrutiny.

🔸 Architecture for agents Where existing components got in the way, the team sometimes rebuilt them, deliberately trading reuse for faster agent-driven development. This is an architectural trade-off worth discussing separately: how much do we gain now, and what will maintenance cost later?

🔸 Tools for observation The agent confused screens and guessed coordinates. The team gave it a way to read the view hierarchy and locate the intended screen. The interface between the agent and its environment matters here: it can check where it is before acting. Context becomes accessible through a tool (this recalls the ACI discussion in our SWE-Agent episode).

🔸 Prototypes alongside production A shared codebase and design system, with experimental code physically excluded from the production build. In my notes, I connected this to discovery and delivery: designers can try ideas, and the result uses the same components as the product. This setup narrows the gap between a prototype and its implementation; the decision about release readiness still rests with an engineer.

Overall, there is little that is new in the article, but what it does contain fits well with the mainstream ideas we discussed in earlier articles.

#AI4SDLC #Agents #Engineering #Architecture #DevTools