Agents & tools
Human in the Loop
An AI workflow that pauses for a human to review, approve, or correct the model’s output at key steps — instead of running fully autonomously.
01 ——
In plain English
Human-in-the-Loop (HITL) is the design pattern of inserting human approval gates inside an AI workflow. The model proposes, the human disposes — at least for the actions that matter most (sending an email, executing a transaction, deleting data).
Why it's the default for production agents:
- Catches hallucinations before they cause real-world damage
- Builds trust with users and regulators
- Captures training data — every correction is a signal you can learn from
- Reduces liability — a human in the loop is often legally meaningful
Common HITL checkpoints:
- Approval before action — agent proposes, human clicks "approve"
- Edit before send — agent drafts, human revises
- Sampling review — agent acts on 99%, human reviews 1% as quality control
- Confidence-based escalation — only the low-confidence cases reach a human
The opposite end: "Human-on-the-loop" — the human supervises a fleet of autonomous agents rather than approving each action. "Human-out-of-the-loop" — full autonomy, used only in low-stakes or reversible contexts.
02 ——
Related terms
Autonomous Agent
An AI agent that operates with minimal human oversight — making and executing decisions independently across long time horizons.
AI Agent
Software that uses an LLM to plan and act — picking tools, taking actions, and adapting based on results to complete a user’s goal.
Agentic Workflow
A multi-step task where an AI agent autonomously decides the steps, uses tools as needed, and works toward a goal with minimal human steering.
Guardrails
Rules and filters that constrain what an AI model can output — used to block harmful, off-topic, or non-compliant responses.
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
AI Evaluation
The structured process of measuring how well an AI model performs — accuracy, safety, cost, latency — usually with a fixed test set called an eval.
Back to glossaryLast reviewed May 2026