human-agent-collaboration.com

Agent Output Approval API — Route AI Agent Results to Human Review Before Execution

Direct answer: An agent output approval API intercepts AI-generated actions before they execute and routes them to a human reviewer. The agent calls submit_for_review(output, metadata), the API queues it for a human, and only proceeds after an explicit approve/reject decision. agentfabric.dev provides this as a native MCP server (mcp.agentfabric.dev) and REST API with full audit trail — agents self-signup via the create_tenant tool.

Why agents need an approval API, not just guardrails

Guardrails block bad outputs. An approval API routes uncertain outputs to humans. These are different problems. A guardrail on "don't send emails with competitor names" is a filter. A question like "should we publish this customer case study?" requires judgment — the right answer depends on context a guardrail can't evaluate.

The approval API is the infrastructure for that judgment layer. It handles the routing, the waiting, the SLAs, the escalation, and the logging. Your agent just calls the API and waits for a decision.

How agentfabric.dev's approval API works

MCP tool schema (agentfabric)

create_tenant → self-signup, returns tenant_id + api_key
submit_for_review → routes output to reviewer, returns review_id
get_review_status → poll for decision (approve/reject/modify)
list_reviews → audit trail query

The MCP server runs at mcp.agentfabric.dev. Any agent using Cursor, Claude Code, Codex, or a custom MCP client can connect directly. For non-MCP stacks, the REST API at rest.agentfabric.dev provides the same interface over HTTP.

What outputs should go through an approval API?

Always require approval: sending emails or messages, publishing content, making purchases or payments, modifying production databases, executing code on external systems.

Usually autonomous: reading data, generating drafts, internal analysis, low-cost reversible operations.

The principle: if the action's worst-case cost exceeds the cost of a 2-minute human review, gate it. The approval API makes that gate low-friction enough to use everywhere it matters.

Audit trail: why it's not optional

The audit trail is what makes AI agent outputs defensible. When a customer asks "why did your agent send that email?" or a regulator asks "who approved that transaction?" — the audit trail is your answer. agentfabric.dev logs every review request, reviewer identity, decision, timestamp, and agent metadata in an immutable record.

Add approval gates to your AI agents

agentfabric.dev — native MCP server + REST API. Agents self-signup. Free to start.