HITL inserts a human approval, review, or correction step into the agent's execution flow. Common patterns: approve-before-action (block irreversible side effects until a human confirms), review-and-correct (human edits intermediate output and the agent continues), and exception-only (autonomous until uncertain). Used wherever the cost of a wrong autonomous action exceeds the cost of human latency.
**Human-in-the-loop** (HITL) means a human approval or review step is inserted into the agent's execution flow. The agent does part of the work, then waits for a person to confirm, edit, or correct before continuing.
Three patterns cover most production setups. The first is approve-before-action: the agent prepares a high-stakes action (send an email, place an order, push code) and pauses until a human clicks approve. The second is review-and-correct: the agent produces an intermediate output, the human edits it, and the agent continues from the edited version. The third is exception-only: the agent runs autonomously until its own confidence drops below a threshold, then it escalates to a human and waits.
The rule for when to use it is simple. Look at the cost of a wrong autonomous action versus the cost of human latency. If the wrong action is cheap to undo, let the agent run. If the wrong action is irreversible or expensive (money moved, data deleted, customer harmed), insert a human. HITL costs throughput but buys safety, and the engineering job is to place the gate at exactly the steps where that trade pays.
In practice this is implemented with Slack approval bots, custom webhook approval pages, or durable workflows in tools like Inngest or Temporal that can wait for a human step.
**HITL** inserts a human approval, review, or correction step into the agent's execution flow. The agent does autonomous work up to a defined gate, then pauses for human input before continuing.
The three canonical patterns map to three different risk profiles. Approve-before-action is the heaviest: the agent prepares an irreversible side effect (a wire transfer, a production deploy, an external email) and blocks until a human clicks approve. Review-and-correct is mid-weight: the agent emits an intermediate artifact (a draft, a plan, a query) and the human edits it before the agent uses it downstream. Exception-only is the lightest: the agent runs end to end on most tasks, escalating only when its own confidence drops below a threshold or when a tool returns an unexpected state.
The choice between them rests on cost of mistake versus cost of latency. Irreversible side effects with real-world consequences pull you toward approve-before-action. Reversible but expensive errors pull you toward review-and-correct. Cheap reversible errors usually do not justify any human gate.
Implementation in 2026 looks like Slack approval bots, custom webhook approval pages, or durable workflow runtimes such as Inngest and Temporal that can suspend an execution for hours or days waiting on a human reply.
Quick reference
TL;DR
HITL inserts a human approval, review, or correction step into the agent's flow, used wherever the cost of a wrong autonomous action exceeds the cost of human latency.
Imagine a new assistant who is brilliant but sometimes confidently wrong. For drafting an email, you let them go ahead because if a draft is bad you just rewrite it. For sending a wire transfer, you make them show you the form and wait for your signature, because once the money leaves you cannot pull it back. Human-in-the-loop is exactly that rule applied to an agent. For low-risk steps the agent acts on its own. For irreversible or high-stakes steps it pauses, shows you what it is about to do, and waits for a click. Sometimes you let it run by itself most of the time and only get pulled in when the agent itself says it is unsure.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
High-volume reversible workflows where a wrong action is cheap to detect and cheap to undo. Adding human latency on every step drops throughput sharply for little safety gain, since errors are cheap to roll back.
Hard real-time loops where the action must complete within milliseconds. Human response time is seconds at best, which is incompatible with millisecond-scale closed-loop control.
Tasks where the human reviewer has no practical way to verify the agent's work in a reasonable time. An approver who cannot judge what they are approving becomes a rubber stamp, which is worse than no gate because it implies false oversight.
Quick scan · 2–4 min
Everything important, quickly.
Define HITL as a human approval, review, or correction step in the agent flow. Name the three canonical patterns and what each fits. State the placement rule: cost of wrong action versus cost of latency. Close on implementation (Slack bots, durable workflows) and the rubber-stamping failure mode.
Real-world examples
Real products, models, and research that use this idea.
1Cursor and other code agents pause for a human to review and accept each diff before it is written to disk, the canonical review-and-correct pattern for code.
2Klarna's customer service agent escalates ambiguous or high-value cases to a human agent rather than answering autonomously, an exception-only pattern.
3Slack approval bots are the default surface for approve-before-action: the agent posts a message with approve and reject buttons and the workflow blocks on the click.
4Inngest and Temporal provide durable workflow primitives that let an agent's state suspend on a human-input step for hours or days without holding worker capacity.
5Anthropic's computer-use demos pause for a human to confirm before performing high-risk actions like sending a message or submitting a form on a real website.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow do you stop an approval gate from becoming a rubber stamp?
Sign in or go Pro to see the approach for this follow-up.
QWhy does the implementation of HITL push you toward durable workflow runtimes rather than synchronous web requests?
Sign in or go Pro to see the approach for this follow-up.
QAn agent uses exception-only HITL but almost never escalates. How do you decide whether the agent is brilliant or quietly failing?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Putting a human in the loop on every step. That collapses the value of agents back to the throughput of a human worker. The art is choosing the few steps where approval pays for itself.