Human-in-the-loop
Deliberately keeping a person at the decision points of automated work — approvals, escalations, and final accountability — while agents do the labor around them.
01What it means
Human-in-the-loop (HITL) is the design pattern where automated work pauses for a person at defined points: approving a change, confirming an action with consequences, answering for the result. It is the alternative both to manual work and to full autonomy — agents own the mechanical layer, humans own the judgment layer, and the boundary is written down rather than improvised.
Our own autonomy policy draws the line concretely. Agents fully own scaffolding, test suites, boilerplate, and first-pass code reading. Humans always own architecture and data modeling, security-relevant changes, anything touching production data or customer systems, and trade-off decisions. And one line does not move: no change reaches a client repository without a human reviewer who understands it well enough to be accountable for it — because accountability is the product a client is paying for when things break at 2am.
Always-on agents sharpen the pattern into policy. Before you hand an agent your tools and let it work around the clock, write down what it may decide alone — thresholds, caps, escalation paths — and what interrupts a human. The wake-you list is policy; everything else is configuration. If you cannot state the boundary, the agent cannot respect it.
02Why it matters for buyers
- —It is what makes AI work answerable: a person who can honestly say "I understand this change" is what you are actually buying when you buy production software from anyone, human or agent-assisted.
- —Specifics beat intentions: ask a vendor which decisions a human makes — by name, with thresholds. "A human reviews everything" and "our AI is safe" are the same non-answer.
- —It is not the slow lane: HITL at decision points with agents doing everything mechanical is faster than the manual status quo — the pauses are minutes, the leverage is constant.
03Common questions
Does human-in-the-loop slow delivery down?
At the decision point, yes — minutes. Everywhere else the agent runs unblocked, so the overall cycle is faster than manual work with the review built in. The slow version of HITL is reviewing everything; the fast version is reviewing what matters.
Which decisions should a human keep?
Anything with consequences that outlive the task: architecture and data modeling, security-relevant changes, production data, and trade-off calls about scope, debt, and what "done" means. The mechanical layer — scaffolding, tests, boilerplate — does not need the pause.
How does HITL relate to agent autonomy?
Autonomy is a dial set per task, and HITL is where you set it. The working rule: delegate tasks with checkable outcomes first, and write down what the agent may decide alone before it runs unattended.