Gatehouse

Governed agents for branded-merch order ops — boundaries enforced at the platform layer, not in the prompt.
aws · us-east-1 · live
agent harness · runtime quoting-agent supplier-agent artwork-agent tool calls (MCP) policy gateway policy.yaml deterministic · not an agent human approver NEEDS_APPROVAL sign-off resumes the call ALLOW DENY → error to agent the call never reaches the tool MCP TOOL SERVERS catalog · decoration pricing supplier stock · POs art proofs · brand specs every decision, allowed or not audit trace DynamoDB observability contract control plane · oversight
Agents never see the policy. Every tool call crosses the gateway; the gateway decides, logs, and only then forwards — enforcement and observability are the same hop.

Scenarios

The agents don't know the policy exists. Nothing here is prompt-enforced — the gateway matches every tool call against policy.yaml and writes the decision before the agent hears back.

policy.yaml — the wall

quoting-agent: allow: [read:catalog, read:decoration_pricing] deny: [read:client_pii] # data boundary place_po: max_usd: 5000 # authority boundary over_limit: require_approval   supplier-agent: allow: [check:supplier_stock, read:lead_times] deny: [place_po] # scope boundary   artwork-agent: allow: [submit:art_proof, read:brand_specs]   default: deny # everything else is a wall

Audit trace

0 decisions
TimeAgentTool callDecisionMatched rule
Run a scenario — every gateway decision lands here, allowed or not.