Agent Task Phase Monitoring Prompt

This prompt is designed for a specific architectural role: an external observer that validates an autonomous agent's self-reported progress against its actual behavior trace. The core job-to-be-done is to detect phase drift—situations where an agent's internal state claims it is on step three, but its recent actions and observations suggest it is stuck on step two, has skipped ahead, or has silently entered an error-recovery loop. The ideal user is an AI engineer or platform developer building a multi-agent system who needs a separate, non-planning monitoring step to trigger replanning, human escalation, or state correction. You must provide the declared plan, the agent's recent action-observation history, and the agent's self-reported current phase. Without all three inputs, the prompt cannot perform its comparative function.

Use this prompt when you have already instrumented your agent to log its tool calls, environmental observations, and internal state updates. The prompt acts as a phase auditor, not a planner. It assumes the plan exists and is well-formed. It does not decide what the agent should do next; it only produces a structured assessment of where the agent actually is relative to the plan. This is critical for systems where agents can hallucinate progress, such as when a web navigation agent reports a form submission as complete despite a CAPTCHA blocking the action. The prompt's output—a phase label, completed milestones, blocked status, and a replanning signal—should be consumed by an orchestrator or a supervisor agent, not shown directly to end users.

Do not use this prompt as a replacement for deterministic state machines when the plan is linear and the environment is fully observable. If your agent's task is a simple sequence of API calls with clear success/failure responses, a code-based state tracker is cheaper, faster, and more reliable. This prompt is valuable when the environment is messy: web pages that change structure, API responses that require interpretation, or multi-step reasoning where success is ambiguous. Also avoid using this prompt in latency-sensitive inner loops. Phase monitoring is a periodic audit, not a per-step check. Run it every N steps, on detected anomalies, or when the agent's self-reported confidence drops below a threshold. The prompt's value is in catching systemic drift, not micro-optimizing step-by-step execution.

Before integrating this prompt, ensure you have a ground-truth mechanism to validate its assessments. The prompt can be wrong—it can misclassify a legitimate workaround as off-plan behavior or miss a subtle phase transition. Implement eval harnesses that compare the prompt's phase detection against human-annotated agent traces. Track false positive replanning triggers (which waste compute and annoy users) and false negatives (which allow agents to continue down broken paths). In high-stakes domains like financial transactions or healthcare actions, always route the prompt's replanning signal through a human approval step before the agent's plan is actually modified.

This template is designed to be invoked at each agent reasoning step or after every tool-call cycle. It compares the agent's current trajectory against the ground-truth execution plan to produce a structured phase assessment. Replace every square-bracket placeholder with live data from your agent's execution context before sending the request to the model.

textCopy
You are an agent task phase monitor. Your job is to compare the agent's current execution state against the original plan and produce a structured phase assessment.

## PLAN
[PLAN]

## COMPLETED MILESTONES
[COMPLETED_MILESTONES]

## AGENT'S LAST ACTION
[LAST_ACTION]

## AGENT'S LAST OBSERVATION
[LAST_OBSERVATION]

## TOOL CALL HISTORY (last 5 calls)
[TOOL_CALL_HISTORY]

## CONSTRAINTS
- Do not hallucinate progress. Only mark a milestone as complete if there is explicit evidence in the observation or tool call history.
- If the agent has repeated the same failing action 3 or more times without progress, flag it as BLOCKED.
- If the agent's last action is unrelated to the current expected phase, flag a possible DIVERGENCE.

## OUTPUT SCHEMA
Return a single JSON object with the following fields:
{
  "current_phase": "string (the name of the current phase from the plan)",
  "phase_status": "IN_PROGRESS | COMPLETED | BLOCKED | NOT_STARTED",
  "completed_milestones": ["string"],
  "next_milestone": "string",
  "is_stalled": true | false,
  "stall_reason": "string or null",
  "is_diverged_from_plan": true | false,
  "divergence_detail": "string or null",
  "replanning_recommended": true | false,
  "replanning_reason": "string or null",
  "assessment_confidence": 0.0-1.0
}

To adapt this template, start by replacing [PLAN] with the full, ordered list of phases and milestones the agent is expected to follow. Populate [COMPLETED_MILESTONES] from your execution log's verified completion records, not from the agent's self-reported claims. The [LAST_ACTION] and [LAST_OBSERVATION] fields should be the raw text of the most recent tool call and its result. If your agent uses a multi-step reasoning loop, invoke this prompt after every observation to catch stalls early. For high-stakes autonomous workflows, route any output where replanning_recommended is true to a human-in-the-loop review queue before the agent takes further action.

The Agent Task Phase Monitoring Prompt is designed to run as a supervisory sidecar within an agent execution loop. It should not be the agent's primary reasoning prompt. Instead, call it at defined checkpoints: after each tool call, when the agent pauses, or when a step fails. The prompt consumes the agent's current plan, completed milestones, recent observations, and the last few action-observation pairs. It returns a structured phase assessment that your harness uses to decide whether to continue, replan, or escalate.

Validation is mandatory. Parse the model's JSON output and validate it against the ground-truth plan state your agent runtime already tracks. Check that completed_milestones are a subset of the plan's defined milestones and that the current_phase label exists in your phase taxonomy. If the model claims a milestone is complete but your execution log shows the associated tool call failed, flag a phase hallucination and discard the assessment. For high-stakes workflows, route conflicting assessments to a human review queue with the agent's trace attached.

Staleness detection is the primary failure mode. An agent can drift from its plan without the phase monitor noticing. Implement a divergence check: compare the agent's last N actions against the expected actions for the reported phase. If the agent is executing tool calls unrelated to the current phase for more than two consecutive steps, mark the phase assessment as stale and force a replanning cycle. Log staleness events with the phase monitor's output, the divergent actions, and the timestamp for post-mortem analysis.

Model choice and latency budget. This prompt is a classification-plus-summary task, not a generation task. Use a fast, instruction-tuned model (e.g., GPT-4o-mini, Claude Haiku, or a fine-tuned small model) rather than a large reasoning model. The prompt should run in parallel with the agent's next action when possible, not block the main loop. Set a strict timeout (500ms–2s depending on your latency requirements) and fall back to the last known valid phase if the monitor times out. Cache the plan and phase taxonomy in the system prompt prefix to reduce per-call token costs.

Wiring the output into the agent loop. On receiving a valid phase assessment, your harness should take three actions: (1) update the agent's visible context with the current phase and blocked status, (2) if replanning_needed is true, inject a replanning instruction into the agent's next turn with the blocked_reason as context, and (3) append the assessment to the session audit log. Never silently discard a blocked status. If the phase monitor reports blocked three consecutive times without resolution, escalate to a human operator with the full plan trace and the last three assessments.