MAAT — the kill switch for runaway AI agents

AI agents don't crash. They loop. A broken tool call doesn't throw; it retries the same prompt, with the same fat context, at full API price, at 2 a.m., on your key. Spend meters like LiteLLM or Helicone tell you what you spent, after the fact. Nobody watches agent behavior at runtime and acts. MAAT is that judgment layer: a drop-in gateway (adoption is one line, the base_url) that watches every workflow. Progress-aware loop detection kills runaway retry loops without killing healthy multi-step agents. a growing conversation is a working agent; only stagnant repetition is a loop. Per-workflow budgets enforce graduated consequences: warn, then downgrade, then kill. Enforcement lives outside the model's context, so a prompt-injected agent cannot talk its way past it. Observe mode runs the same judgments in shadow to tune thresholds on production traffic risk-free, and every kill produces a one-click incident report with the measured burn rate projected forward. Built on AMD end to end: the gateway runs live on an AMD Developer Cloud MI300X droplet; the primary tier is Fireworks AI serverless (Kimi K2.6 at real prices); and near-budget workflows degrade to Gemma 3 4B served by vLLM on the same MI300X via ROCm — AMD-hosted Gemma at $0, so agents degrade instead of dying. Verified on real traffic: a retry loop carrying ~4.9k tokens per call was killed on its third identical attempt — $0.0124 burned in 8.3 seconds, a measured pace of ~$322 over an unattended weekend. Gateway overhead is ~7 ms p50, and the full test suite (guard logic plus HTTP integration) passes. Try it live during the judging window: the dashboard at http://129.212.191.62:8080 is public, and the README includes a deliberately public, budget-capped demo key. a leaked key bounded by MAAT costs at most its budget. That is the product.