FORGE — AI Software Engineer

Problem. Debugging, validating, and releasing software is where developer time leaks. A bug hides across files: reproducing it, tracing the root cause, fixing it, re-running tests, and clearing a release gate can eat hours of manual, error-prone work. Non-engineers can't meaningfully participate at all. Solution — FORGE on IBM Bob 2.0. You describe a problem in plain words. FORGE maps the repository, runs the real test suite, and orchestrates specialized Bob subagents in parallel — architect, debugger, tester, actor, critic, release. Each is a Bob task with real codebase context: it reads files, edits code, and returns structured verification. Every claim is an evidence node (file:line, call chain, diff, test output) so nothing is fabricated. The critic and release gate only pass on real, green test evidence. Target users & interaction. Anyone — founders, PMs, QA, and students — types the problem and watches repository → root cause → fix → tests → security → release gate; engineers get an audit trail and a one-click PR. Creative & unique. Honest by construction: the system refuses to release unverified changes. Deterministic skill learning: every verified Bob-backed fix is distilled into a replayable skill, so repeat defects are fixed with zero LLM calls. Fully observable: every change is a diff, every step an evidence graph. Effectiveness. In end-to-end runs it reproduced and fixed a real checkout race condition on a sample project: failing regression tests → suite green → critic 7/7 → release-ready, with an audit trail and PR.