Promptetheus

A full stack AI agent observability and action platform. Promptetheus captures a production agent run, detects the likely failure, replays the exact bad step, and ships a fix with regression evidence. It is built for teams running LLM agents who need to debug failures they cannot reproduce.

AI agents fail in ways that are hard to debug. They click the wrong button, ignore a user correction, retrieve the right policy and then answer the opposite, or claim success without doing the work. Logs alone do not catch these, because the run reports success while the outcome is wrong. Promptetheus is a full stack observability and action platform that closes the loop from failure to fix instead of stopping at a log viewer. It instruments an agent run, classifies likely failures (goal mismatch, false success, policy contradiction, ignored warnings), replays the exact session with an event timeline and screen and video evidence, attributes the bad outcome to a single critical step, and then turns that incident into a fix path. The fix path can be a fix brief, a patch target, a real GitHub PR, or a deterministic fallback. It then re-runs the original bad step as a regression replay and records before and after evidence so the fix is proven rather than assumed. The architecture is a real hosted service. FastAPI is the single write gateway for all trace-derived state, Supabase (Postgres, Auth, Storage, RLS) is canonical storage with workspace isolation, and a Next.js console renders replay, incidents, logs, docs, and the guided demo over an authenticated SSE stream.