Marshal · Humans in the agent loop
HITL Approvals
Human-in-the-loop approvals, inputs, and selections—agents pause mid-run for a human decision and resume exactly where they left off
Marshal · HITL Approvals
How HITL Approvals works
Human-in-the-loop approvals, inputs, and selections—agents pause mid-run for a human decision and resume exactly where they left off
HITL
ask · decide · resume
Agents open requests (sky rails); humans resolve them and the SDK unblocks the run (orange rails).
Agents open requests (sky rails); humans resolve them and the SDK unblocks the run (orange rails).
Why it matters
Built for production agents—not prompt playgrounds
How this harness capability fits the Exemplar platform: one identity, one policy layer, and one audit trail across console, MCP, and SDK.
Why Exemplar
Full autonomy is the wrong default for production: deploys, migrations, and spend need a human at the decision point, not after the incident.
HITL makes escalation a first-class harness primitive—no webhook servers, callback queues, or bespoke approval UIs per agent.
What Exemplar delivers
Blocking SDK primitive—harness.hitl().ask(...)—that works from LangGraph, CrewAI, Agno, ADK, or any framework that calls Python.
One console inbox where reviewers approve, reject, answer, or pick an option, with agent-supplied context and full audit on every decision.
How teams use it
Gate high-impact actions—production deploys, credential rotation, tier-0 changes—behind an approval the agent requests itself.
Wrap ask(...) as a framework tool so the model escalates to a human when confidence is low; the run resumes with the human's answer.
Capability checklist
Developer guide
Official documentation on docs.exemplar.dev for this capability.
Open developer guide (opens in a new tab)Get started
Put HITL Approvals to work with the rest of Marshal—or talk to us about your agent stack.
From the blog
Related posts on exemplar.dev.
- AI Agent Governance: How to Control AI Agents Running in Production
AI agent governance is the set of policies, controls, and audit mechanisms that determine what AI agents can do, when they need human approval, and how their actions are logged. The five pillars, how governance differs from the harness, and why it matters for compliance.
- The Harness Engineering Checklist
15 things to put in place before trusting AI-generated code in production — organised by phase: foundation, enforcement, task design, and maintenance. The checklist most teams wish they had before they started.
- Agents, context, and guardrails on a unified platform
From code completion to production actions; Context Lake, catalog, governance, and DevX Assist/MCP for safe automation.