Chor

Gatekeeper: a procurement agent whose guardrails evolve themselves (Statement 1: Recursive Harnessing) AI agents are starting to approve vendors, and they fail in two ways. They miss sanctioned companies hidden behind misspellings and subsidiaries (OFAC's 50% rule), and they "de-risk" by nationality: our baseline agent wrongly blocked 44% of legitimate vendors. Gatekeeper keeps the model fixed and evolves the harness around it. The harness is a versioned policy "genome" in MongoDB Atlas. It controls which tools the agent may use, which checks must run before approval, fuzzy-matching and ownership-graph depth, and evidence-bound reject/escalate rules. A deterministic Strands hook enforces it, so the model can't override it. A critic (Claude Sonnet 5) reads the agent's failures, recalls past attempts and what they broke through Atlas Vector Search, and proposes structured policy changes. Candidates are scored against hidden labels and promoted only if held-out catch never drops. Atlas change streams hot-swap each new champion into every running agent in about 100–250 ms. It runs on real data: 16,610 US-sanctioned parties via Atlas Search (fuzzy), and the GLEIF ownership graph via $graphLookup. Results (same model, held-out cases, 3 runs each): - Catch rate: 87% → 100% - Legitimate vendors wrongly blocked: 44% → 3.7% - Prompt-injection catch: 80% → 100% - Cost: about $0.02 per vendor check Built entirely in the Atlas Hackathon Sandbox.

Team

Placement

Participant