Incident2Fix

The problem: production incidents are toil. An alert fires at 3am; an engineer reads it, digs through logs, reproduces the bug, writes a fix, then runs tests hoping nothing else broke. It is repetitive, stressful, high-stakes work — exactly the kind of work AI agents should take on. But no engineering team will let an agent push code unsupervised, so most "AI incident tools" stop at suggestions in a chat window and never touch the code. Our solution: Incident2Fix, a supervised agentic pipeline that turns a production incident report into a verified, tested code fix — with humans holding the two decisions that matter. Four IBM Bob 2.0 agents work the case in sequence: a Triage Agent confirms severity and ranks suspect files; a Root-Cause Agent reproduces the bug and isolates the faulty code with evidence; a Fix Agent proposes a minimal diff; a Regression Test Agent generates tests pinning the incident and runs the full suite. Between diagnosis and fix sits Human Gate 1 (approve the root-cause analysis before any fix is generated); between the proposed diff and the codebase sits Human Gate 2 (approve the patch before anything is applied). Reject either gate and the pipeline halts with zero changes. Every stage writes a structured report, and a Flask dashboard renders the whole evidence trail — incident context, RCA, the diff in green/red, gate decisions, test results — so each run is fully auditable. Target users are on-call engineers, SREs, and small teams without dedicated incident-response staff. They interact with it three ways: drop an incident report (markdown) into the pipeline, answer y/n at the two approval gates, and review the outcome in the dashboard. A full run is one command plus two keystrokes of approval. What makes it creative and unique: human approval is a first-class pipeline stage rather than a bolt-on, and it closes the loop other tools leave open — from alert all the way to merged, tested code. Nothing is hardcoded: the orchestrator never contains a diagnosis; Bob reasons over the incident, the code, and prior reports on every run. Efficiency: three seeded incidents (a EUR checkout 500, a silent discount bug, an inventory corruption) each ran end-to-end on two human keystrokes, growing the suite from 5 baseline tests to 45 passing, with 15 evidence reports preserved. The toil compresses to minutes of human attention per incident.