BobFlow — Proof, Not Prompts

AI coding assistants have collapsed software development into a single gesture: describe what you want, receive code, and hope it works. The bottleneck simply moved. Generation became cheap; trust did not. A developer still has to read the diff, guess at the edge cases, and discover the security holes in production. For teams this is worse than it sounds, because AI-generated code arrives without the reasoning that produced it — review becomes archaeology. The problem is not that AI writes bad code. It is that AI writes code without evidence. BobFlow is an evidence-driven delivery workflow for IBM Bob 2.0. Instead of one prompt producing one patch, work moves through nine explicit phases: Understand, Plan, Investigation, Synthesize, Implement, Verify, Fix/Retest, Review, Delivery. The distinctive stage is phase three. Three read-only specialist investigations — Architecture, Security, and Testing — examine the same feature request, each with its own mandate and output contract. None of them may edit a file. Their findings are reconciled in a synthesis step that must resolve conflicts explicitly rather than silently averaging them. Only two of the nine phases are permitted to modify code. Everything before is investigation; everything after is verification. BobFlow is built for software teams adopting AI agents in codebases where correctness and security are non-negotiable — platform teams, regulated industries, and any group that has been burned by a confident-sounding generated patch. They interact with it the way they already work: they hand IBM Bob a feature request, not a prompt. What comes back is not just code but the reasoning, the threat model, the test matrix, and the raw output of the test run that proves it. Most agent frameworks optimise for autonomy — fewer humans, more machine decisions. BobFlow optimises for auditability. Its design accepts a deliberate cost: read-only investigation phases are slower than letting an agent start writing immediately. In exchange, the delivery artifact is something a reviewer can actually check. The verification phase is the point. BobFlow never reports a test result it did not observe. In our demo, the evidence is the raw terminal output, not a summary written afterwards. We believe the next competitive axis for AI coding tools is not how fast they generate, but how much of their output a human can trust without re-deriving it. BobFlow is a concrete answer to that question.