Release Captain

The Problem Every software team faces the same moment before a release: is this safe to ship? Today, that usually means someone manually: 1) Reads recent commits and guesses their risk 2) Checks whether new dependencies were added 3) Runs the test suite and reads the output 4) Checks whether requirements are actually covered by tests 5) Makes a judgment call, often under time pressure We timed this by hand on a real project — a TypeScript to-do list API with an existing automated test suite. A clean release took 14 minutes 36 seconds; one with a real regression took 12 minutes 22 seconds, because even a broken release needs the same steps to find out it's broken. Worse, the only way we caught the regression was by actually running the tests — the commit message ("no functional changes intended") gave no warning at all. The Solution: Release Captain Release Captain automates this judgment call using four independent AI subagents, each built with IBM Bob 2.0: 1) Changelog — summarizes what actually changed since the last release 2) Risk — flags dependency and configuration risk 3) Test — runs the existing test suite and reports pass/fail 4) Spec Coverage — checks whether stated requirements have matching code and tests An orchestrator runs all four in parallel and applies one rule: any red flag from any subagent means NO-GO. No single check has to be perfect — the whole verdict only needs one subagent to catch a real problem. Target Users & Interaction Release managers, tech leads, and small teams without dedicated QA who need a fast, defensible go/no-go call. Users interact through a simple web report (built with Streamlit): pick a branch, get a verdict, drill into any subagent's findings. What Makes It Unique 1) Built recursively: we used Bob to build a tool that judges whether software built the way Bob builds it is safe to ship — a fitting use of the technology on itself. 2) No single point of failure: while testing, we found a real blind spot — our Spec Coverage subagent can be fooled by code that merely looks correct. We didn't hide this. We used it to prove the design works: even when one subagent is wrong, the Test subagent still catches the real failure independently, so the final verdict is right anyway. 3) Proven on a real bug, not a toy example: our test project has an actual security regression — deleting a to-do item silently stopped requiring an API key — and Release Captain correctly calls NO-GO on it. Effectiveness & Efficiency A manual check takes over 12 minutes and depends on someone remembering every step. Release Captain runs the same checks automatically, every time, with nothing skipped under deadline pressure — turning a slow, memory-dependent human process into a fast, repeatable, evidence-backed decision.