Human-in-the-Loop Handoff State Prompt

This prompt is designed for agent workflows that pause execution at a predefined approval gate and must hand a concise, actionable summary to a human reviewer. The reviewer needs to understand what the agent has done, what decision is required now, what evidence supports the proposed action, what risks are present, and what happens if they approve or reject. Use this prompt when the cost of a wrong autonomous decision exceeds the cost of human review time—for example, before sending a high-value email, executing a database migration, approving a financial transaction, or publishing content that carries legal or reputational risk. The prompt assumes the agent has already executed prior steps, collected tool outputs, and reached a point where it cannot proceed without explicit human authorization.

Do not use this prompt for real-time streaming decisions where latency makes human review impractical, trivial approvals where the agent's confidence is already calibrated above your defined threshold, or situations where the human reviewer already has full context from another source such as a dashboard or notification system. This prompt is also inappropriate when the approval decision requires domain expertise the reviewer does not possess—handing a complex medical or legal decision to a non-expert reviewer creates a false sense of safety. Before deploying this prompt, verify that your reviewer pool has the necessary context, authority, and training to evaluate the agent's proposed action. The prompt's value comes from compressing agent state into a reviewer-optimized format, not from replacing the reviewer's judgment.

The ideal user of this prompt is an engineering team building agent orchestration systems that include explicit human approval nodes. You should already have infrastructure for pausing agent execution, queuing review tasks, capturing reviewer decisions, and resuming or aborting the workflow based on the response. This prompt is the interface between your agent runtime and your human operations layer. To implement it correctly, you must define the approval schema upfront: what decision types are possible (approve, reject, modify, escalate), what evidence must be presented for each decision type, and what metadata your logging and audit systems require. The prompt template in the next section provides the structure; your application code must enforce that the agent cannot proceed until a valid reviewer response is received and validated against your schema.

Next, copy the prompt template and adapt the placeholders to your specific approval gate. Pay particular attention to the [RISK_LEVEL] and [CONSTRAINTS] placeholders—these directly shape how the agent communicates uncertainty and what guardrails it applies to its own recommendations. After adapting the template, build your eval suite using the criteria in the Evaluation and Testing section to ensure the handoff summaries are complete, accurate, and decision-ready before you put a human reviewer in the loop.

This prompt template is designed to be pasted into your agent's handoff module. It takes the agent's current execution state—including the active plan, completed steps, pending decisions, tool outputs, and any risk flags—and produces a structured summary that a human reviewer can act on without reading the full trace. The template uses square-bracket placeholders that you wire to your agent runtime's state object before calling the model.

textCopy
You are an agent handoff summarizer. Your job is to produce a concise, actionable summary of the current agent state for a human reviewer who must approve or reject pending decisions.

## Current Agent State
- Active Plan: [ACTIVE_PLAN]
- Completed Steps: [COMPLETED_STEPS]
- Current Step: [CURRENT_STEP]
- Pending Decisions: [PENDING_DECISIONS]
- Tool Outputs (last 5): [RECENT_TOOL_OUTPUTS]
- Risk Flags: [RISK_FLAGS]
- Assumptions Made: [ASSUMPTIONS]
- Confidence Scores: [CONFIDENCE_SCORES]

## Output Schema
Return a JSON object with the following fields:
{
  "summary": "One-paragraph summary of what the agent is doing and why.",
  "pending_decisions": [
    {
      "id": "decision_id",
      "description": "What needs to be decided.",
      "options": ["option_a", "option_b"],
      "agent_recommendation": "Recommended option with rationale.",
      "risk_level": "low|medium|high",
      "consequences_if_wrong": "What happens if this decision is wrong."
    }
  ],
  "evidence_summary": "Key evidence the agent used to reach this point.",
  "risk_assessment": {
    "overall_risk": "low|medium|high",
    "top_risks": ["risk_1", "risk_2"],
    "irreversible_actions_taken": ["action_1"]
  },
  "next_steps_if_approved": ["step_1", "step_2"],
  "questions_for_reviewer": ["question_1"],
  "missing_information": ["info_gap_1"]
}

## Constraints
- Do not omit the `consequences_if_wrong` field for any pending decision.
- If no irreversible actions have been taken, return an empty array, not null.
- Flag any decision where the agent's confidence is below [CONFIDENCE_THRESHOLD].
- If the agent has made assumptions, list them explicitly in the summary.
- Keep the summary under 200 words.
- Use the exact risk levels provided; do not invent new categories.

To adapt this template, replace each bracketed placeholder with values from your agent's execution state. The [ACTIVE_PLAN] should include the full plan structure with dependencies. [PENDING_DECISIONS] must capture every gate where the agent is waiting for human input. Set [CONFIDENCE_THRESHOLD] to a value appropriate for your risk tolerance—typically 0.7 for high-stakes workflows, 0.5 for exploratory ones. Before deploying, validate that your state serialization produces valid JSON for each placeholder and that the output schema matches your review queue's expected format. For high-risk domains, always include a human-readable fallback rendering alongside the structured JSON.

The Human-in-the-Loop Handoff State Prompt is not a standalone artifact—it is a contract between the agent runtime and the human review queue. Wire this prompt into your application at the exact point where the agent pauses execution and requires a human decision. The prompt consumes the agent's current state snapshot (plan progress, pending decisions, tool outputs, risk flags) and produces a structured reviewer-ready brief. The output must be validated before it reaches a human reviewer, because an incomplete or misleading handoff summary causes decision errors that compound across the rest of the workflow.

Integrate the prompt into an API call that passes the agent's serialized state as [AGENT_STATE], the list of pending decisions as [PENDING_DECISIONS], and any risk classification as [RISK_LEVEL]. On the response path, validate the output against the expected [OUTPUT_SCHEMA] before enqueueing it for human review. Required fields include decision_summary, action_context, risk_flags, and expected_outcomes. If validation fails, retry once with the validation errors injected into [CONSTRAINTS] as explicit missing-field or malformed-output instructions. After a second failure, log the raw response and the validation errors, then escalate to an on-call channel rather than silently retrying. For high-risk domains (healthcare, finance, legal), always route the validated handoff through a human-review queue that records the reviewer's decision, timestamp, and any override rationale. This creates an audit trail that links the agent's state at pause time to the human action taken.

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 or older models that may drop required fields or hallucinate risk flags. If your agent runtime uses tool-calling, include the tool call history and outputs in [AGENT_STATE] so the handoff summary accurately reflects what the agent has already attempted. For logging, capture the prompt version, the agent state hash, the validated handoff output, the reviewer's decision, and the time-to-decision latency. These logs become essential for debugging decision errors, auditing compliance, and measuring whether the handoff prompt is producing actionable summaries or causing review fatigue. Do not treat this prompt as a fire-and-forget text generation step—treat it as a state serialization and decision interface that must be tested, monitored, and versioned alongside the rest of your agent harness.