Agent-to-Human Handoff Summary Prompt Template

This prompt template is designed for operations teams and platform engineers who need to bridge the gap between autonomous AI agents and human reviewers. Its primary job is to transform raw agent execution traces, tool-call logs, and internal state into a structured, immediately actionable summary for a human operator. The ideal user is a developer or AI architect integrating an agent into a support, moderation, or review workflow where the agent handles routine tasks but must escalate complex, ambiguous, or high-risk cases. The required context includes the agent's original objective, the actions it took, the evidence it gathered, and the specific decision or input now required from the human.

Use this prompt when the cost of a wrong autonomous decision is high, when the agent has exhausted its capability or authority, or when a workflow policy mandates human approval. It is specifically designed for scenarios where the human reviewer is time-constrained and must not read raw logs. Do not use this prompt for simple status updates that don't require a decision, for logging agent activity to a passive audit trail, or for handoffs between two autonomous agents. For agent-to-agent handoffs, use a structured state-transfer object, not a natural-language summary designed for human cognition. This prompt is also inappropriate when the agent's task was trivial and a handoff would create more overhead than value.

Before implementing, define the human's decision space clearly. The prompt must constrain the output to present only the information a human needs to make that specific decision. Avoid the common failure mode of dumping all agent context into the summary. Instead, the harness should filter the agent's trace to include only actions relevant to the pending decision, evidence that supports or contradicts the agent's recommendation, and a clear statement of what the human is being asked to do. The next step is to pair this prompt with an evaluation rubric that checks for completeness, clarity, and the absence of hallucinated actions or evidence.

This template produces a structured handoff summary that a human reviewer can act on without reading raw agent logs. It forces the model to separate what the agent did from what the human needs to decide, and it requires explicit next-step labeling. Use this template when an agent reaches a decision boundary, exhausts its authority, hits a confidence threshold, or encounters an exception that requires human judgment. Do not use it for routine status updates or internal agent-to-agent communication—those need different schemas with less decision framing.

textCopy
You are an agent preparing a handoff summary for a human reviewer. Your goal is to provide exactly the context the human needs to make a decision or take action—no more, no less.

Generate a handoff summary using the structure below. Use square-bracket placeholders only where data is unavailable; otherwise, provide the actual value.

## HANDOFF SUMMARY

### Task Context
- **Original Request:** [RESTATE_THE_USER_REQUEST_OR_TRIGGERING_EVENT]
- **Task ID:** [TASK_IDENTIFIER]
- **Agent Role:** [YOUR_ROLE_AND_CAPABILITY_BOUNDARY]
- **Handoff Reason:** [WHY_THIS_REQUIRES_HUMAN_INPUT: e.g., confidence below threshold, action outside authority, exception encountered, approval required]

### Actions Taken
[List each action you performed, one per line, with outcome]
- [ACTION_1] → [OUTCOME_1]
- [ACTION_2] → [OUTCOME_2]

### Evidence Collected
[For each piece of evidence, include source and relevance]
- **Source:** [SOURCE_IDENTIFIER_OR_REFERENCE]
  - **Excerpt:** [RELEVANT_EXCERPT]
  - **Relevance:** [WHY_THIS_MATTERS_TO_THE_DECISION]

### Current State
- **Completed:** [WHAT_IS_DONE]
- **Pending:** [WHAT_REMAINS]
- **Blockers:** [ANY_OBSTACLES_PREVENTING_PROGRESS]

### Decision Required
- **Question for Human:** [THE_SPECIFIC_QUESTION_YOU_NEED_ANSWERED]
- **Options Presented:** [LIST_2-5_OPTIONS_WITH_TRADE_OFFS]
- **Recommendation:** [YOUR_RECOMMENDED_OPTION_WITH_RATIONALE]
- **Deadline/Urgency:** [TIME_SENSITIVITY_IF_ANY]

### Confidence Assessment
- **Overall Confidence:** [HIGH/MEDIUM/LOW]
- **Key Uncertainties:** [WHAT_YOU_ARE_UNSURE_ABOUT]
- **Worst-Case If Wrong:** [POTENTIAL_IMPACT_OF_INCORRECT_ACTION]

### Next Steps After Human Decision
- **If Approved:** [WHAT_HAPPENS_NEXT]
- **If Rejected:** [ALTERNATE_PATH]
- **If More Info Needed:** [WHAT_CLARIFICATION_IS_REQUIRED]

[CONSTRAINTS]
- Do not omit any section. If a section has no content, write "None" or "N/A" with a brief explanation.
- Cite evidence sources explicitly. Do not summarize evidence without linking it to a source.
- If confidence is LOW, the recommendation must include explicit uncertainty language.
- Keep the summary under [MAX_WORD_COUNT] words. Prioritize decision-relevant information.
- Use plain language. Avoid agent-internal jargon that a human reviewer would not understand.

Adaptation guidance: Replace [MAX_WORD_COUNT] with a concrete limit appropriate to your review interface—typically 300–500 words for a quick review, 800–1,200 for complex decisions. The [CONSTRAINTS] block is where you add domain-specific rules, such as regulatory disclosure requirements, tone policies, or mandatory fields. If your handoff feeds into a ticketing system, add a [TICKET_FIELDS] placeholder for system-mandated metadata. For high-risk domains like healthcare or finance, add a constraint requiring explicit flagging of any action that could cause harm if the human decision is delayed.

Before deploying, test this template against at least three handoff scenarios: a routine approval, an edge-case exception, and a low-confidence escalation. Validate that the output includes all required sections, that evidence citations are traceable to source data, and that the decision question is specific enough for a human to answer without follow-up. If your eval shows that recommendations are consistently overconfident or evidence is thin, tighten the confidence-assessment instructions or add a requirement to list missing evidence explicitly.

The Agent-to-Human Handoff Summary Prompt is designed to sit at the boundary between an autonomous agent's execution loop and a human review queue. In a typical implementation, the agent calls this prompt as its final action when it determines that human input is required—either because it has completed its work and needs approval, has encountered an ambiguity it cannot resolve, or has hit a capability boundary. The prompt consumes the agent's internal state (actions taken, evidence gathered, pending decisions) and produces a structured handoff payload that a downstream review system can parse, display, and route. The harness is responsible for ensuring that the handoff is complete, actionable, and traceable before it reaches a human reviewer.

Wire this prompt into your application by wrapping it in a function that accepts the agent's execution trace as input. Before calling the model, assemble the [AGENT_ACTIONS], [EVIDENCE_COLLECTED], [PENDING_DECISIONS], and [REQUIRED_HUMAN_INPUT] fields from the agent's internal state—do not pass raw logs. After receiving the model's output, validate the JSON schema strictly: confirm that required_human_action is a non-empty string, that pending_decisions is an array with at least one item, and that each decision includes a context and options field. If validation fails, retry once with the validation error message appended to the prompt as a correction hint. If the retry also fails, log the failure, attach the raw agent trace as a fallback, and route to a human with a flag indicating that the summary may be incomplete. For high-risk domains such as healthcare or compliance, always require a human to confirm the summary against the raw trace before acting on it.

Model choice matters here. Use a model with strong instruction-following and structured output capabilities—GPT-4o, Claude 3.5 Sonnet, or equivalent. Avoid smaller models that may drop required fields or hallucinate actions the agent did not take. If your agent system already uses tool calling, consider implementing the handoff as a structured tool response rather than a free-text prompt: define a handoff_to_human function with the summary fields as parameters, and have the agent call it directly. This approach gives you schema validation at the tool-calling layer and reduces the risk of malformed JSON. Log every handoff with a unique handoff_id, the agent's session identifier, a timestamp, and the full prompt and response for auditability. Route handoffs to human reviewers based on the priority and category fields in the output—high-priority items should bypass the normal queue and trigger an alert. Finally, capture the human's decision (approved, rejected, modified, escalated) and feed it back into your evaluation pipeline to measure whether handoff summaries are driving correct and timely human actions.