WiredFront

VibeGuard is a self-tightening agent harness built for vibecoders — people who ship with AI coding tools and pay for every retry, runaway loop and leaked key. The agent attempts coding tasks, a deterministic judge scores them, and every failure category on a training task becomes a proposed guardrail rule carrying its provenance (the failing attempt, the matched snippet, the detector). Only a human can approve — or loosen — a rule; the next run loads approved rules automatically. Held-out pass rate climbed 1/6 → 3/6 → 5/6 → 6/6 over five runs with zero false blocks, with Wilson 95% intervals on every point. All state lives in MongoDB Atlas as the Agent Engine’s four memory types — procedural (versioned rules), episodic (run traces), semantic (vibecoder lexicon with Voyage embeddings and Atlas Vector Search), taxonomic (failure categories) — plus an append-only audit of receipts, and is mirrored into Atlas Agent Engine hosted memory, whose build_context() hands the approved rules back as the agent’s context policy. A COSTAR side-conversation scores each prompt and emits a form of at most two questions before any money is spent. Deployed on Modelence Cloud; Raindrop-instrumented (proposals as AI events, human verdicts as signals). Statement 1: recursive harnessing. Statement 2: long-horizon memory driven by a hard metric. Agents tighten their own leash; only humans loosen it.