Sentinel Swarm

Right now, AI coding tools only fix bugs a human already found. They wait for a ticket. Sentinel Swarm doesn't wait — it goes looking. We send four IBM Bob 2.0 subagents into a codebase with no bug report and no hints. Scout reads the code and figures out what each function is supposed to do. Saboteur then writes a real test targeting the edge case most likely to break it — empty input, a missing field, a boundary number. We actually run that test. If it fails, that's not a guess dressed up as an insight — it's proof, on screen, in real time. Medic then writes a fix and re-runs the exact same test to confirm it now passes. Scribe writes up what happened in plain English: what broke, why, how it was fixed, and how severe it was. That's the whole idea in one sentence: most tools suggest bugs, we prove them. Target users are developers and engineering teams shipping code faster than they can review it — exactly the gap current research (the Active-SWE benchmark) shows AI agents are weakest at: proactively finding problems nobody asked about. Sentinel Swarm turns that weakness into the entire product. A judge doesn't need to install anything or trust our word for it. They visit the live app, pick a scenario (or paste in their own small Python snippet — safely checked and sandboxed before anything runs), and click one button: "Release the Swarm." Four panels light up in sequence as the real chain runs — Scout mapping, Saboteur attacking, a test visibly failing, Medic patching, the same test turning green, and a case-file card assembling itself with a severity score and a "proven, not guessed" badge. The whole thing takes under two minutes and needs nothing from the judge but a click. What makes this different from existing AI code review tools isn't cleverness — it's a rule we never break: nothing is reported as a bug unless we've watched it fail with our own eyes. That single constraint is also our safeguard against hallucination — if the generated test doesn't actually fail, there's no finding, no report, no false claim. Efficiency-wise, this replaces a process that today requires a human to notice something's wrong, write it up, wait for someone to pick it up, and only then get a fix — often after the bug's already shipped. Sentinel Swarm collapses that entire cycle into one autonomous, self-verifying loop, running end to end in under two minutes with zero human bug report required. Built entirely on IBM Bob 2.0's Agent mode and subagent architecture — every "smart" decision in the pipeline is a real Bob call, not hardcoded logic pretending to be one.