Put humans in the loop without breaking the pipeline.
When an agent reaches a decision it cannot make alone, it surfaces a structured request and waits. The pipeline does not break — it pauses.
Three steps. No broken pipelines.
Agent reaches a decision boundary
The HitlAgent evaluates the current state and determines it cannot proceed autonomously — a budget threshold, a legal grey area, an infrastructure change above a risk tier. It surfaces a structured escalation request and suspends execution.
Structured escalation surfaces in the HITL inbox
The request appears in the human operator's inbox with full context: the agent identity, the current step, the decision it cannot make, and the options available. No truncated logs. No ambiguous state.
Human approves, rejects, or redirects — pipeline resumes
The operator responds. The HitlAgent receives a typed resolution, updates state, and continues execution from the exact point it paused. No restart. No replay from scratch. The pipeline resumes with the decision recorded.
Built for the decisions that matter.
Budget approval
Any expenditure above a defined threshold requires human sign-off before the financial intent executes.
Infrastructure changes
Scale-down operations, region migrations, or config changes on production systems pause for explicit operator approval.
Legal review
Contracts, compliance decisions, or any action flagged by the governance engine for legal sensitivity route through HITL before execution.
Risk decisions
Agents operating in regulated environments can declare risk tiers. Actions above the autonomous threshold surface for human review automatically.
Not a failure state. A design guarantee.
In most agent frameworks, human escalation is a workaround — a try/catch that routes to a Slack channel and hopes someone sees it. In Agiorcx, HitlAgent is a first-class SDK primitive. You declare escalation thresholds the same way you declare state transitions.
The difference matters at audit time. When a regulator asks why a particular decision was made, the answer is not "the agent timed out and a human intervened." The answer is "the agent evaluated the decision boundary, surfaced the request at step 4 of 7, the named operator approved it at 14:22:08Z, and execution resumed at step 5." Full attribution. Full traceability.