Artificial Intelligence

Meta Engineers Forum App Using AI-Native Development

Meta has shared insights into how engineers built its newest standalone application, Forum, by integrating an AI agent directly into the development cycle.

Radwan AbdulhadiEditor and technology writer

Published: October 15, 2026 · Updated: October 15, 2026 · 3 min read

Meta Engineers Forum App Using AI-Native Development — Image: Meta Engineering

Building Forum With Agentic AI

Meta engineers adopted a completely new way of working while building the company's newest application, Forum, designed for users wanting to engage deeply with conversations in their Facebook groups. Detailed in an official update shared via Meta Engineering, the standalone Groups app was constructed by a small team from scratch using an agentic AI framework.

Rather than utilizing artificial intelligence merely as a faster typing assistant, the team treated the AI agent as an active participant in the engineering process. According to the development team, success was not measured by code volume, but by verified and production-ready outcomes that successfully built, passed tests, and survived code review.

Shifting From Human-in-the-Loop to Human-on-the-Loop

The project shifted the ownership of each development step. Under a traditional human-in-the-loop setup, the engineer drives every step while the AI assists internally by autocompleting functions or suggesting fixes. In contrast, Meta experimented with a human-on-the-loop model where the agent executes entire iterations—planning, building, testing, correcting, and repeating—while the engineer sets goals, monitors runs, steps in during drift, and approves results.

This operational shift allowed cycles to overlap rather than running strictly sequentially, shifting engineers' time toward direction and judgment. However, the team noted that this approach relies entirely on trusted verification. Without automated validation, AI-generated code risks producing plausible yet flawed outputs that require extensive debugging.

Automated Verification and Technical Specifications

To ensure reliability, Meta set the completion bar at verification rather than simple compilation. Agents checked their own work using automated pass and fail signals, unit tests, and UI specification matching before presenting pull requests to human engineers. This process ensured that what reached review was a smaller set of high-confidence results complete with build results and recorded verification sessions.

Four main capabilities powered this autonomous loop: rigorous upstream decomposition for planning specifications, automated verification inside the loop, simulated user testing with navigation agents interacting with the live app, and agentic production operations that assist with post-launch root-cause analysis after user feedback.

Machinery and Architectural Legibility

The operational machinery relied on five key components: permanent project rules, loadable skills capturing team conventions, automated checks, commands chaining workflows together, and automated hooks acting as triggers for linters and tests. Furthermore, engineers prioritized architectural legibility over inherited complexity, keeping the codebase straightforward and modularizing critical components so agents could easily plug and play.

To resolve issues where agents lost track of application states, engineers provided the AI with direct tools to read live view hierarchies from simulators rather than relying on textual descriptions. Additionally, the team utilized a separate build flavor as a sandbox application, sharing the exact production codebase while keeping early-stage prototypes isolated from final builds.