Eswar Dev

ForgeGate is a self-extending ops agent where no generated tool runs until it passes an independent, five-stage verification gate. When an ops task has no matching tool in the registry, a coding model writes a Python CLI tool; that tool is then run through (0) a fail-closed AST policy, (1) behavioral cases in isolated subprocesses, (2) a semantic review by a judge model, (3) adversarial probing by a verifier model, and (4) a determinism replay. Only a clean verdict promotes the tool into permanent procedural memory, where it is reused and its usage counted. A stage-2 pass with a stage-3 break is a recorded "disagreement" and fails closed — a tool is never promoted because one model liked it. The whole thing is standard library only (dependencies = []), every model call is raw urllib + json, and the offline path is the default so the benchmark never silently needs a funded key. 25 tasks across 8 tool types and three modalities (text, voice, screenshot) run end to end; the gate rejects four deliberately broken tools at four different stages.