Human-in-the-loop in an agentic workflow isn't an emergency fallback — it's a moment designed in advance, where a decision returns to a human with a defined threshold, a specific owner, and the full case context, before the agent ever starts running.

Three modes of human involvement

Three recurring patterns show up in practice, differing in exactly when the human steps in:

  • Approve-before — the agent prepares a decision but doesn't execute it until a human clicks "approve." Used where a mistake is costly or hard to reverse, such as issuing a refund above a certain amount.
  • Monitor-after — the agent acts on its own, and a human reviews a sample of decisions periodically, not in real time. Suits reversible, low-stakes decisions, like standard replies to order-status questions.
  • Escalate-on-exception — the agent acts on its own for routine cases and only hands off when it hits the edge signal. This is the dominant mode across most agentic workflows — the one shown in the diagram on the Agent Flow homepage.

How to pick the approval threshold

Choosing the mode and threshold is really a trade-off between two costs: the cost of a wrong decision the agent makes on its own, and the cost of a human's time spent approving it. The more expensive and harder to reverse a decision is, the closer you want to sit to approve-before. The cheaper and more reversible it is, the more monitor-after or plain escalate-on-exception is justified. A good starting point is estimating what the worst plausible mistake in a given decision category would actually cost — and weighing that against the time cost of a human approving every single case individually.

In practice the threshold is rarely a single number — it's often a combination of amount and context, e.g. "refunds up to $80 auto-approved, above that approve-before, except for VIP accounts, where every amount goes to a human." The more conditions like that exist, the more it matters that they live in one written place rather than scattered across a few people's heads.

What a human should get at the handoff

A bare alert saying "human decision required," with no context, just shifts back onto the human exactly the work the agent was supposed to do — reading the case from scratch. An effective handoff includes: the full case history, a one-sentence reason the agent couldn't decide on its own, and — if the agent has a tentative recommendation — that recommendation along with its confidence level. That's what turns approval time from minutes into seconds.

Human-in-the-loop in framework practice

The mechanism for pausing a process and routing it for approval is now part of the core architecture of major agent frameworks, not a bolt-on. LangGraph documents it as an interrupt, where workflow state is persisted and execution only resumes after a human decision. That direction lines up with broader guidance on human-AI interaction design — the Guidelines for Human-AI Interaction (Amershi et al., CHI 2019, Microsoft Research) recommend, among other things, that a system clearly signal uncertainty and make it cheap and fast to correct its decisions. That's exactly what a well-designed handoff point in an agentic workflow does.

The risk of approving too much or too little

Too wide an approval threshold leads to decision fatigue — a human clicks "approve" without reading, because they see dozens of similar cases a day, and the whole mechanism stops doing its job. Too narrow a threshold creates a bottleneck instead: the agent waits on approvals that are, in practice, always positive, and the process ends up slower than before automation. Both extremes are worth checking after the fact — if the share of approvals with no changes exceeds roughly 95%, the threshold is probably too cautious.

Connecting back to ticket triage

The same choice of mode shows up in the ticket-triage example: standard order-status questions run under escalate-on-exception, while a refund above a set amount needs approve-before. We walk through designing that workflow step by step in Designing an Agent Workflow for Ticket Handling, and cover the concept of an agentic workflow itself in What Is an Agentic Workflow?.

Who should approve — one person or a team

With approve-before, it's worth deciding upfront whether the decision sits with one specific person (a department lead, say) or with anyone holding a given role on a given shift. A single decision-owner speeds things up but creates a bottleneck the moment they're out — which is why, in practice, a role tends to work better than a name: "whoever's on duty in the complaints queue," not "Sam from accounting." That seemingly minor call often determines whether the workflow survives the first time a key person takes a vacation.

Does human-in-the-loop slow the process down compared to full automation? For a single case — yes, minutes instead of seconds. For the process as a whole — usually not, because it only touches exceptions, not every case. If it turns out to touch most cases, that's a sign the agent's decision boundaries are set too narrow, not that the approval mechanism itself is unnecessary.