The Human-Agent Handoff Problem

Aviation has spent decades engineering the handoff between autopilot and human pilot. The Airbus A330 flight deck, the Boeing 777, every modern commercial aircraft embodies lessons learned from incidents where the transition of control between automatic systems and human crew went wrong. The literature on this is substantial and sobering: mode confusion (the pilot doesn’t know whether the autopilot is engaged and in what mode), automation surprise (the aircraft does something the pilot didn’t expect and they react wrongly), skill erosion from underflying (pilots who rarely manually fly the aircraft lose proficiency when they need it), and handoff at the wrong moment (transferring control when the situation is most complex rather than most straightforward).

Human factors engineers have names for all of these failure patterns, protocols designed to prevent them, and required crew training that specifically addresses the automation-to-human transition. This body of knowledge took decades and cost lives to develop.

The agentic AI industry is going to re-derive these lessons from first principles, at enterprise scale, without the benefit of the aviation industry’s institutional memory. The human-agent handoff problem is not the same as the aviation handoff problem — the stakes and failure modes differ — but it is structurally similar enough that ignoring the analogy is a choice that will cost money.

The handoff in agentic systems occurs in both directions. The “agent to human” handoff happens when the agent reaches a decision point it cannot resolve, an error condition it cannot recover from, or a task completion that requires human review. The “human to agent” handoff happens when a human delegates a task to an agent, transferring responsibility and providing the context the agent needs to proceed. Both directions have characteristic failure modes that are underappreciated and poorly designed around.

The agent-to-human handoff fails most often through a combination of timing and context. Timing: the agent escalates at the moment it encounters an obstacle, which is often not the moment when human attention is available or most valuable. Context: the agent hands off insufficient information about what it has done, what it was trying to do, what it found uncertain, and what the human needs to decide. A human receiving a handoff from an agent typically sees: the immediate question the agent is stuck on, but not the chain of reasoning that led there or the alternatives the agent considered.

This is the equivalent of handing a complex legal matter to a new paralegal mid-task by dropping a single document on their desk with a sticky note saying “What should we do about this?” without providing the case history. The paralegal might be able to handle the specific question. They cannot assess whether the question they’re being asked is the right question, or whether the path that led to this point was sensible.

The context problem at handoff is tractable but rarely addressed. An agent capable of producing a structured handoff summary — what was the original task, what steps have been taken, what information was gathered, what decision is being escalated and why, what are the options the agent identified — gives the receiving human the material they need to make an informed decision. This requires that the agent be explicitly designed to produce this summary rather than simply stopping with a question.

Most current agent systems are not designed this way. They are designed to escalate when they get stuck, not to prepare a comprehensive handoff. The result is that human reviewers spend significant time reconstructing context from logs (if logs are comprehensive enough to support this reconstruction, which is not guaranteed) or, more commonly, simply restarting the task from the beginning to ensure they understand where they are.

The wasted time is measurable. In a process that involves five agent-to-human escalations, each requiring twenty minutes of context reconstruction, the escalation overhead is nearly two hours of human time that produces no output — it only restores a starting point. Designing for clean handoffs with structured context would reduce this to perhaps thirty minutes total. The engineering investment required is modest; the operational return is significant.

The human-to-agent handoff fails differently. The primary failure mode is instruction underspecification: the human provides a task description that is clear to another human but ambiguous to the agent in ways the human did not anticipate. The agent encounters the ambiguity mid-task, makes an interpretive choice, and proceeds on that interpretation without surfacing it to the human. The human later receives an output that addresses a somewhat different question than they intended.

This failure is so common that practitioners have started describing it as the “telephone game” of agent delegation: the agent interprets the instruction through its own priors and training, produces output faithful to its interpretation, which may be a plausible but incorrect reading of what the human actually wanted. The gap between intended and interpreted is often small enough that the output looks plausible until examined closely.

The defense requires that agents be designed to surface interpretive choices, particularly early in task execution when correction is cheap. “You asked me to prepare a competitive analysis. I’m going to focus on the three largest competitors by revenue. Before I proceed: is that the right scope?” This kind of confirmation step adds friction to the workflow. It also dramatically reduces the frequency of large misalignments between human intent and agent execution that produce expensive re-work. The tradeoff is almost always worth making for tasks above a certain complexity threshold; the challenge is designing agents that make it automatically rather than leaving it to the operator to remember to ask.

The mode confusion problem from aviation appears in agentic systems as a different but analogous issue: humans interacting with agent-assisted workflows who are uncertain whether a given output or decision came from an agent or a human, and who treat the information differently (or identically) depending on the answer. This matters because human-generated information and agent-generated information warrant different scrutiny, different trust calibration, and different escalation thresholds.

In practice, the provenance of outputs in agent-assisted workflows is often obscured rather than surfaced. A document management system that uses agents to extract and categorize information presents extracted data to human reviewers without indicating which data was agent-extracted and which was human-entered. The reviewer, unable to distinguish, applies the same (often insufficient) scrutiny to both. Agent-generated errors slip through review processes designed for human-generated inputs.

Provenance tagging — explicitly marking every piece of information with whether it was generated by a human or an agent, and if an agent, which agent operating under what configuration — is a practice that several regulatory-forward organizations have implemented and that should become standard. The implementation is not technically complex. It requires discipline in system design and process documentation that many organizations do not maintain.

There is a longer-horizon dimension to the human-agent handoff problem that goes beyond the operational issues. As agents handle more of the routine, lower-judgment tasks in knowledge work, the humans in agent-assisted workflows are increasingly doing only the escalations, reviews, and exceptions — the unusual and difficult cases. This changes the cognitive profile of the human role in ways that create risks analogous to the aviation skill erosion problem.

A claims adjuster who previously reviewed a hundred claims per week and exercised judgment on all of them, but now reviews ten agent-processed claims per week for quality assurance and handles twenty escalated exceptions, is developing a different skill set than the previous version of the job. The judgment on unusual cases may remain sharp because the escalated cases are, by definition, unusual. The ability to process routine cases quickly may atrophy because the routine cases are no longer in the workflow. If the agent system fails — if the organization needs to revert to manual processing for any reason — the human capacity to absorb the restored workload may not be there.

This is not an argument against agent deployment. It is an argument for maintaining human engagement with the full range of tasks in some minimum fraction, ensuring the skills that may be needed in contingency scenarios don’t erode entirely. Aviation requires simulator training that includes manual flight scenarios precisely because autopilot usage erodes manual proficiency. Enterprise agent deployments do not yet have equivalent training requirements, and the organizations that will most miss them are the ones that haven’t thought about it until the contingency arises.

The handoff is not an edge case in agentic system design. It is the joint — the place where two systems that operate differently connect. Joints, as any structural engineer knows, are where things break.

The quality of the human-to-agent handoff is partly a design problem and partly an information problem. The agent needs to know not just what task to do but what context frames the task — what the human was trying to achieve, what constraints apply, what prior work is relevant, and what success looks like in terms the human cares about (which may differ from the metrics the agent was trained to optimize for). Encoding all of this in a task description is harder than it appears; natural language is ambiguous in exactly the ways that generate mismatches between human intent and agent execution.

Several teams building agent-assisted workflows have addressed this by developing structured handoff templates — not free-form task descriptions but structured forms that require the human to specify goal, constraints, success criteria, out-of-scope items, and escalation triggers before handing work to the agent. The structure sacrifices some flexibility but dramatically reduces the interpretive latitude that leads to misalignment. The templates are essentially requirements specifications applied to agent task delegation — a practice that software engineering figured out decades ago and that knowledge work is reinventing for the agent context.

The organizational implication of taking handoff quality seriously is that the human role in an agent-assisted workflow becomes more demanding in some ways even as it becomes less demanding in others. The human is doing less of the routine execution work. They are doing more of the work of specifying, reviewing, and course-correcting — which requires a different kind of expertise and attention. Training programs that help knowledge workers develop the specific skills required to work effectively alongside agents — task specification, critical output review, exception judgment — are beginning to appear, but they are not yet standard in organizations deploying agents at scale.

The organizations that invest in this training — that treat “how to work effectively with agents” as a legitimate professional skill requiring deliberate development — will have a structural advantage over those that assume the handoff is intuitive. It is not intuitive. It is a learned capability, and the people who develop it produce dramatically better outcomes from their agent-assisted work than those who treat agents as interchangeable with the colleagues they partially replaced.