Meta's Muse: the prototype took 2 weeks, reliability took 7 months

Alexandr Wang, Meta's Chief AI Officer, says Muse went from a 1–2 week prototype inspired by OpenClaw to launch 7 months later. The team tracked more than 100 agent behaviors, built evals for them, and set a launch-blocking threshold for each row; checkpoints that passed 80 of the rows did not ship. Why it matters: Wang's account separates demo speed from launch readiness: the team tuned the model and product together, used explicit behavioral gates, and treated trust as something earned through progressively larger tasks. This is a Meta executive's description of its process, not independent evidence that the process made Muse reliable.

Try this: List 20–50 behaviors your agent must get right, attach an eval and a launch-blocking threshold to each, then record every red row in the release decision.