111 — Agent-to-Human Handoffs

Concept overview

A handoff is a workflow transition in which responsibility moves from an agent to an accountable person or team. It includes the current state, completed actions, unresolved questions, relevant evidence, and a clear requested decision. Displaying a vague escalation message is insufficient if no recipient accepts the work.

Separate a request for approval from a transfer of operational ownership. Some cases need a single yes-or-no action decision; others require investigation, judgment, or continued customer contact. Define who owns the case while waiting, when the request expires, how changed evidence invalidates it, and what an approved resumption may do.

Why it matters in production

Proof of concept

A proof of concept can pause for one reviewer and inspect whether the evidence packet supports a concrete decision without rereading the whole conversation.

Production

Production requires authenticated queues, acceptance tracking, deadline escalation, version-bound decisions, reassignment rules, and explicit agent behavior while the case waits.

Enterprise scale

Enterprise scale adds workload balancing, skill-based routing, accessibility, shift transitions, reviewer quality checks, and metrics separating agent preparation from human queue delay.

Architecture components

Handoff trigger policy

Detects authority limits, uncertainty, conflicts, deadlines, and situations requiring professional judgment.

Decision packet builder

Presents verified facts, sources, completed effects, unresolved items, and the exact requested action.

Human work queue

Routes to qualified recipients, tracks acceptance, and maintains a named owner during waiting and reassignment.

Decision endpoint

Authenticates reviewer responses and binds them to the current evidence and proposed action revision.

Resumption controller

Rechecks preconditions, permissions, and decision scope before allowing further agent work.

The architecture, mapped

sequenceDiagram
    participant A as "Agent workflow"
    participant Q as "Human work queue"
    participant H as "Qualified reviewer"
    participant R as "Resumption policy"
    A->>Q: Submit versioned decision packet
    Q->>H: Assign and request acceptance
    H->>Q: Accept ownership
    H->>R: Submit scoped decision
    R->>A: Validate current evidence and authority
    A-->>R: Context unchanged
    R->>A: Resume approved action only

Runtime flow

1. 1

   Detect need for transfer

   Evaluate explicit policy and task state rather than relying on the model's willingness to ask for help.

2. 2

   Freeze decision context

   Record evidence versions, completed effects, current uncertainty, and the proposed action or investigation question.

3. 3

   Route to accountable recipient

   Select a qualified queue and retain current ownership until a person or team accepts the case.

4. 4

   Support informed review

   Provide concise evidence with access to originals, exact consequences, alternatives, and a way to request clarification.

5. 5

   Validate the decision

   Authenticate the reviewer, check authority and expiry, and reject decisions bound to superseded evidence or action parameters.

6. 6

   Resume or transfer fully

   Execute only the approved scope, or record that the human now owns ongoing work; reconcile effects and close remaining obligations explicitly.

Referral ambiguity handoff

Healthcare · Illustrative enterprise scenario, not a reported customer outcome.

Business situation

Illustrative architecture: A referral coordinator's assistant finds conflicting patient identifiers and incomplete administrative documents.

The problem

A generic request to consult staff may leave the referral unowned while the agent continues retrieving the wrong chart.

Agent requirement

Stop identity-dependent retrieval and transfer a minimal evidence packet to the authorized registration team with explicit ownership.

Architecture

The workflow enters a waiting state, submits a scoped identity-resolution packet, and resumes document collection only after an authenticated resolution references the current referral version.

Runtime

• Detect identifier conflict
• stop dependent retrieval
• build minimal packet
• route registration queue
• record acceptance
• validate resolution
• resume scoped collection

Failure scenario

The receiving team does not accept before the service window expires; escalation retains the original coordinator as owner until reassignment is acknowledged.

Operational impact

Measure packet completeness, acceptance time, decision rework, stale-approval rejection, and unresolved-case age; the system supports administration and does not make clinical decisions.

Architecture tradeoffs

Failure scenarios & recovery

Common architecture mistakes

Sending the entire transcript instead of a decision packet.

Summarize verified facts, evidence links, completed effects, and the precise decision needed.

Treating silence as approval.

Keep the workflow paused or transfer through an explicit documented escalation path.

Allowing a human click to authorize arbitrary future actions.

Bind the decision to exact operations, resources, limits, and expiry.

When to choose this pattern

Choose it when

• A workflow reaches authority limits or requires professional judgment.
• Evidence conflicts cannot be resolved safely through bounded automated checks.
• Humans need to own exceptions across shifts, departments, or external organizations.

Avoid it when

• A deterministic input validation error can be returned directly to the requester.
• The proposed queue has no accountable recipient or service expectation.
• An approval prompt merely disguises a decision whose consequences the reviewer cannot inspect.

Architecture checklist

• Define triggers for approval, investigation, and full ownership transfer separately.
• Include evidence, completed actions, uncertainty, and exact requested decisions.
• Track recipient acceptance and preserve ownership during waiting.
• Authenticate reviewers and verify their authority for the proposed action.
• Bind decisions to evidence versions, action parameters, and expiry.
• Specify timeout escalation, reassignment, cancellation, and resumption behavior.
• Measure human queue delay separately from agent preparation and execution.

Related concepts

• 3 — Agentic State Machines
• 12 — State Persistence
• 16 — Task Interruption & Resumption
• 23 — Attribute-Based Access Control (ABAC)
• 30 — Non-Repudiation Logs
• 44 — Human-in-the-Loop (HITL)
• 49 — Timeout Policies
• 58 — Partial Success Management

Primary references

• Temporal: Workflow execution
• NIST AI Risk Management Framework
• NIST SP 800-207: Zero Trust Architecture