Agent Drift Detection and Correction Prompt

This prompt is a diagnostic and corrective tool for agent observability systems that monitor autonomous agent trajectories. Its primary job is to programmatically detect when an agent's current action path has diverged from its stated plan and to produce a structured, machine-readable assessment with a re-aligned next step. The ideal user is an AI reliability engineer or agent-framework builder integrating this prompt into a runtime harness, not an end-user interacting with a chat interface. You should use this prompt when your agent runtime has access to both the original plan and the sequence of completed actions, and you need an automated recovery signal to prevent cascading failures without restarting the entire agent loop.

Concretely, deploy this prompt inside a drift_check function that is called after every N actions or when a heuristic monitor flags a potential divergence. The function must provide the prompt with the original [PLAN], the [ACTION_HISTORY], and the [CURRENT_OBSERVATION]. The expected output is a JSON object containing a drift_detected boolean, a drift_assessment string, and a corrected_next_action object. This structured output allows the agent harness to branch logic: if drift_detected is true, the harness replaces the current next action with the corrected one; if false, it continues. This prompt is not a replacement for a planner. It corrects local drift, not a fundamentally flawed plan. Do not use it for generating initial plans, for agents without a traceable action history, or for evaluating the quality of a final answer to a user.

Avoid using this prompt in isolation without evaluation guards. False-positive drift flags can cause an agent to unnecessarily second-guess a valid path, leading to thrashing. Implement a cooldown or a minimum action threshold before re-invoking the drift check to prevent over-correction. In high-stakes environments where a corrected action could have irreversible side effects (e.g., sending an email, modifying a database record), always route the corrected_next_action to a human approval queue before execution. The next step after integrating this prompt is to build a set of evaluation cases with known drift and non-drift trajectories to calibrate your detection threshold and ensure the prompt's assessment aligns with your operational definition of a plan deviation.

This template is designed to be injected into an agent harness when an observability system flags a potential divergence from the stated plan. It forces the model to compare the original objective and planned trajectory against the actual actions taken, then produce a structured drift assessment and a single, corrected next step. The prompt is self-contained and can be used as a standalone correction call or as a system-level instruction that persists across turns.

textCopy
You are an agent trajectory auditor. Your task is to detect and correct plan drift.

[ORIGINAL_OBJECTIVE]
[OBJECTIVE]

[ORIGINAL_PLAN]
[PLAN_STEPS]

[ACTUAL_TRAJECTORY]
[COMPLETED_ACTIONS]

[DRIFT_THRESHOLD]
[THRESHOLD_DESCRIPTION]

[CONSTRAINTS]
[CONSTRAINTS]

[OUTPUT_SCHEMA]
{
  "drift_detected": boolean,
  "drift_severity": "none" | "minor" | "major" | "critical",
  "drift_evidence": [
    {
      "planned_step": string,
      "actual_action": string,
      "divergence_description": string
    }
  ],
  "goal_still_achievable": boolean,
  "recovery_action": {
    "action_type": "continue" | "correct" | "rollback" | "replan" | "escalate",
    "corrected_next_step": string,
    "rollback_target_step_index": number | null,
    "revised_sub_plan": [string] | null,
    "escalation_reason": string | null
  },
  "confidence": number
}

[EXAMPLES]
[FEW_SHOT_EXAMPLES]

[RISK_LEVEL]
[RISK_LEVEL]

Instructions:
1. Compare each step in [ORIGINAL_PLAN] against the actions in [ACTUAL_TRAJECTORY].
2. If any action deviates from the plan, assess whether the deviation serves the original objective or represents drift.
3. Classify drift severity using [DRIFT_THRESHOLD].
4. If drift is detected, produce exactly one recovery action. Do not propose multiple alternatives.
5. If the goal is no longer achievable, set recovery_action.action_type to "escalate" and provide a clear escalation_reason.
6. If the trajectory is on track, set drift_detected to false and recovery_action.action_type to "continue".
7. Output valid JSON matching [OUTPUT_SCHEMA] exactly. No additional text.

To adapt this template, replace each square-bracket placeholder with runtime values from your agent harness. [ORIGINAL_OBJECTIVE] should contain the user's original request or goal statement. [ORIGINAL_PLAN] should list the planned steps in order. [ACTUAL_TRAJECTORY] should include the sequence of actions the agent actually took, including tool calls, responses, and any errors encountered. [DRIFT_THRESHOLD] defines what constitutes minor, major, and critical drift for your use case—for example, 'minor: action order swapped but goal unaffected; major: skipped a required verification step; critical: goal no longer achievable.' [CONSTRAINTS] should include any hard rules the agent must follow, such as 'never modify production data without confirmation.' [FEW_SHOT_EXAMPLES] should provide 2-3 examples of drift detection and correction to calibrate the model's judgment. [RISK_LEVEL] should be set to 'low', 'medium', 'high', or 'critical' to guide the model's conservatism. After receiving the output, validate the JSON against the schema before acting on the recovery_action. For high-risk or critical workflows, route the drift assessment to a human reviewer before executing any rollback or replan action.

The drift detection prompt is not a standalone artifact; it is a component inside an agent observability harness that compares the agent's current trajectory against the stated plan at configurable intervals. Wire this prompt into your agent runtime as a sidecar evaluation step that fires after every N tool calls, after high-risk actions, or when a trajectory monitor detects semantic divergence from the plan embedding. The harness should supply the prompt with the original plan, the action history, the current state, and any tool outputs that triggered the drift check. The prompt returns a drift assessment and a re-aligned next step, but the harness is responsible for deciding whether to apply that correction, request human review, or escalate.

Build the harness with three guard layers. First, validate the prompt output schema: the drift severity score must be a float between 0.0 and 1.0, the drift_detected field must be a boolean, and the corrected_next_action must include a tool name, arguments, and a justification string. If the output fails schema validation, retry once with a stricter format instruction appended to the prompt. Second, implement a false-positive drift check: if the drift severity exceeds your threshold but the corrected action is identical to the agent's current intended action, suppress the correction and log the event for threshold tuning. Third, enforce an over-correction guard: compare the corrected action's goal alignment score against the original plan's goal embedding. If the corrected action would discard more than 20% of completed progress without clear justification, route to a human reviewer instead of applying the correction automatically. Log every drift check with the plan version, trajectory snapshot, severity score, correction applied, and whether human review was triggered.

For model choice, use a fast instruction-following model with low latency for the drift check path—this prompt runs frequently and must not become a bottleneck. Reserve larger reasoning models for the correction path only when drift severity exceeds 0.7 or when the harness detects a pattern of repeated low-severity drift flags across consecutive checks. Track a drift budget per agent session: if more than three corrections are applied within a single task, stop autonomous correction and escalate the full trajectory to a human operator with the accumulated drift log. Never allow the drift prompt to modify the original plan directly; it may only suggest re-aligned next actions within the existing plan's constraints. The harness must preserve the original plan as the source of truth and append corrections as annotated deviations.