Autonomous SRE agents fail in production for two reasons: a single-shot LLM can't reliably diagnose and fix a real incident, and there's no honest way to know whether the agent caused a recovery or whether Kubernetes just self-healed. We built a harness that solves the first by searching for fixes and grounds the second in a real, decoupled oracle — on a live Kubernetes cluster. The harness wraps a fixed LLM (no model training) and improves how it remediates through bounded tree search: the LLM proposes candidate actions, a synthesized verifier prunes unsafe branches before they execute, safe actions run on the real cluster with undo on abandoned branches, and an out-of-band oracle scores whether the incident actually resolved. The search is the engine of improvement — it turns an agent that fails a cascade single-shot (0/15) into one that resolves it 8/20, by giving it multiple diagnose→act→verify rounds instead of one guess. We adapted the Code World Models / Auto-Harness method (DeepMind) to SRE, where tree search both finds remediations at the incident level and evolves the safety verifier from real outcomes. Three validated results, all on real infrastructure: (1) an attribution oracle with a no-agent control arm that credits the agent only for recovery it caused — attributed delta +1.00 [0.62, 1.00] on persistent faults, and correctly reports no signal when the agent genuinely fails; (2) tree-search auto-remediation that cracks a cascade single-shot couldn't, with an attributable lift; (3) a measured safety gap — the synthesized verifier is accurate (0.87) but lets through 38.5% of dangerous actions vs 15.4% hand-written, the exact metric a deployable harness must close. Every number traces to a committed result file; thin margins, an open-weight-only model sweep, and the still-untested safety coupling are reported, not hidden.
Team
Placement
Participant
