Project 02
Agent output governance for enterprise AI — a transparent MCP middleware that intercepts agent decisions, classifies them by risk, and escalates what matters before execution.
The Problem
AI agents are making procurement decisions, approving invoices, reallocating budgets, and modifying employee records — at machine speed, 24/7. Your team can't review thousands of agent outputs per day.
Companies either slow everything down with manual approval gates on every action, let everything through and discover problems after the damage is done, or build custom governance logic into every agent.
None of these approaches scale with the agent fleet.
The Solution
Resonance Proxy sits on the Model Context Protocol (MCP) — the open standard for AI agent communication. It intercepts agent actions at the protocol level, so you don't need to modify your agents.
Every action gets classified by risk. Low-risk decisions auto-approve and log. Medium-risk items batch into digests for periodic review. High-risk actions escalate immediately — before they execute.
A team of five can govern a fleet of hundreds of agents, reviewing only what matters while maintaining a complete audit trail.
Process
From agent action to governed execution in five transparent steps.
An AI agent issues a tool call over the MCP protocol using Server-Sent Events (SSE). The agent operates normally — no SDK or code changes required.
Resonance Proxy sits between the agent and the target system as MCP middleware. Every tool call is intercepted before execution — transparently and without latency impact.
The classification engine evaluates the action against configurable risk policies: amount thresholds, affected systems, data sensitivity, and custom business rules.
Based on classification, the action routes to one of three paths: auto-approve with audit logging, batch for periodic human review, or escalate for immediate approval.
Approved actions execute normally. Escalated actions are held until a human reviewer approves them through the governance dashboard. Full audit trail maintained.
Risk Classification
Every intercepted action is classified and routed based on configurable risk policies.
Auto-approve & log
Actions that pose minimal risk are automatically approved and logged for audit. No human intervention required.
Examples:
Batch for review
Actions batched into human-digestible digests for periodic review. Teams review on their own schedule.
Examples:
Immediate escalation
Critical actions are escalated immediately for human approval before execution. Zero tolerance for unauthorized high-risk decisions.
Examples:
Why Resonance Proxy
Operates at the MCP protocol level — no agent modifications, no SDK integration, no vendor lock-in.
Drop-in deployment. Your existing MCP-compatible agents work unchanged. The proxy is invisible to them.
Risk thresholds, affected systems, and business rules are configurable per-organization and per-agent fleet.
Complete audit trail of every agent decision — approved, batched, or escalated. Compliance-ready from day one.
Technology
MCP Protocol
Standard for AI agent communication
TypeScript
Core proxy logic and classification engine
Node.js
Runtime and SSE stream handling
Server-Sent Events
Real-time agent-proxy communication
Playwright
End-to-end integration testing
Vitest
Unit and integration test suite
Market Context
The enterprise AI agent market is projected to see 40% of enterprise apps embedding AI agents by end of 2026. As agent autonomy increases, the governance gap widens.
By operating at the MCP protocol level, Resonance Proxy provides a universal governance layer that works across any MCP-compatible agent without modification — regardless of the underlying model or framework.
The batched review model means a team of five can govern a fleet of hundreds of agents, reviewing only what matters while maintaining a complete audit trail of every automated decision.
“Your AI agents make thousands of decisions. How many did you actually approve?”