Trace2Fix is an autonomous incident investigation and remediation system built as a skill for IBM Bob 2.0. Instead of one agent guessing at a fix from an error message, it spawns four independent, read-only investigator sub-agents, each restricted to a single evidence domain: logs, source code, documentation/config, and test coverage. None sees the others' findings — this epistemic isolation prevents anchoring bias, the tendency of a single agent to fixate on its first plausible hypothesis. A parent agent synthesizes the four reports, requiring at least two independent domains to corroborate a claim before it's marked high-confidence, and flags any contradictions. It then prints a structured Root Cause Report and stops — no file is touched. Remediation only begins after a human replies with explicit approval. Post-approval, Bob follows strict TDD: write a regression test and confirm it fails, apply the smallest possible fix, rerun targeted and full test suites, check documentation/config impact, then generate a complete incident report. A 10-point verification checklist must pass fully before anything is marked resolved. We validated this on a seeded FastAPI payment-service outage (INC-001): config/prod.yaml was missing an EXCHANGE_RATE key, causing an unguarded TypeError → HTTP 500 on non-USD payments in production only. The incident file, logs, code, docs, and tests each held one piece of evidence — none the whole answer. All four investigators independently converged on the same root cause. Bob wrote a failing regression test, applied a two-file fix (a None guard raising a clear ValueError, plus a handler returning HTTP 422 instead of a bare 500), left the missing config value alone (a business decision, not a code fix), and verified 10/10 checks. A benchmark CLI comparing manual vs. Trace2Fix runs on INC-001 showed roughly 71% faster time-to-root-cause and 68% lower total MTTR, with correct diagnosis on the first attempt and a regression test manual runs often skipped. The result generalizes beyond this demo: most real incidents sit at the intersection of what logs show, what code does, what config/docs claim, and what tests do or don't cover. Trace2Fix is a template for scoping an agentic on-call assistant narrowly, evidence-first, and unwilling to touch production until a human has read and approved its reasoning.
Category tags: