Agent Plan Dead-End Recovery and Backtracking Prompt

This prompt is for agent-framework builders and AI reliability engineers who need to recover an autonomous agent that has exhausted all viable actions within a single plan branch. The job-to-be-done is not to restart the task from scratch, but to produce a structured backtracking summary that identifies the dead-end, selects the most promising alternative decision point, and proposes a new branch to explore—all while preserving the original goal and the context of what was already tried. The ideal user is someone integrating this into an agent harness where the runtime can detect a dead-end condition (e.g., all tool calls in a branch returned terminal failures, or a sub-planner returned an empty action set) and needs a prompt that turns that signal into a recoverable state.

Use this prompt when your agent architecture supports explicit plan tracking with identifiable decision points and action branches. The required context includes the original goal, the full action history with outcomes, and a representation of the decision tree or plan structure. Do not use this prompt for simple retry loops where a single tool call failed with a correctable argument error—that's a job for a tool-call retry prompt. Do not use it when the agent has no branching logic or when the failure is due to an unrecoverable environmental condition (e.g., total API outage). This prompt assumes the dead-end is a planning failure, not an infrastructure failure. If the agent's state has been corrupted or the goal itself is no longer valid, escalate to a human-in-the-loop workflow instead of attempting automated backtracking.

Before wiring this prompt into production, define clear dead-end detection criteria in your harness: a dead-end is not just one failed action, but a state where all candidate next actions from the current plan node have been attempted and failed, or where the planner returns an empty action set. The harness should also maintain a backtracking budget—a limit on how many times the agent can backtrack before escalating to a human operator. Without this budget, an agent can oscillate between branches indefinitely. The next step after implementing this prompt is to pair it with eval checks that verify the backtracking summary correctly identifies the last viable decision point and that the proposed alternative branch is both distinct from the exhausted branch and feasible given the remaining tool capabilities.

This template is designed for an agent harness that has detected plan exhaustion—every action in the current branch has been attempted and failed, or the branch has been fully explored without reaching the goal. The prompt instructs the model to produce a structured backtracking summary that identifies the dead-end, selects the most recent viable decision point, and proposes a new action from an alternative branch. It is not a generic retry prompt; it assumes the agent has already exhausted local recovery options and must re-anchor to an earlier plan state.

codeCopy
You are an agent execution monitor responsible for recovering from plan dead-ends. Your task is to analyze the current execution trace, identify the point where all available actions in the current branch have been exhausted, and produce a backtracking plan that selects an alternative branch from an earlier decision point.

## INPUT

[CURRENT_PLAN]

[EXECUTION_TRACE]

[TOOL_CATALOG]

[CONSTRAINTS]

## INSTRUCTIONS

1. Examine [EXECUTION_TRACE] and identify the most recent decision point where an alternative action branch was available but not selected.
2. Confirm that all actions in the current branch have been attempted and failed or are no longer viable. If any viable action remains, do not backtrack—instead output a continuation action.
3. Identify the specific decision point in [CURRENT_PLAN] that corresponds to the backtrack target. Reference it by step ID or description.
4. Select an alternative branch from that decision point that has not been attempted. Explain why this branch is the best candidate given [CONSTRAINTS] and the goal.
5. Produce a single next action from the selected branch using a tool from [TOOL_CATALOG]. Include the tool name, arguments, and a brief rationale.
6. Summarize what was learned from the dead-end branch that should inform the new attempt.

## OUTPUT SCHEMA

Return a JSON object with the following fields:
- "dead_end_detected": boolean (true if all actions in the current branch are exhausted)
- "dead_end_branch_id": string (identifier for the exhausted branch)
- "backtrack_target_step_id": string (identifier for the decision point to backtrack to)
- "backtrack_target_description": string (what the decision point was)
- "alternative_branch_id": string (identifier for the selected alternative branch)
- "alternative_branch_rationale": string (why this branch was chosen)
- "next_action": object with fields "tool_name", "arguments" (object), and "rationale" (string)
- "lessons_from_dead_end": array of strings (insights from the failed branch)
- "confidence": string (one of "high", "medium", "low")

## CONSTRAINTS

- Do not propose an action that appears in [EXECUTION_TRACE] as already attempted and failed unless the arguments are materially different.
- Do not backtrack further than necessary. Prefer the most recent viable decision point.
- If no alternative branch exists at any decision point, set "dead_end_detected": true and "backtrack_target_step_id": null, and output an escalation recommendation in "next_action.rationale".
- Preserve the original goal from [CURRENT_PLAN]. Do not change the objective.

Adaptation notes: Replace [CURRENT_PLAN] with the agent's structured plan (steps, branches, decision points with IDs). [EXECUTION_TRACE] should contain the full action-observation history with tool calls, results, and error messages. [TOOL_CATALOG] must list available tools with their schemas so the model can propose valid alternative actions. [CONSTRAINTS] should include budget limits, safety rules, and domain-specific boundaries. For high-risk domains, add a require_human_approval boolean to the output schema and route the backtracking proposal through a review queue before execution. Validate the output against the schema before allowing the agent to act on the proposed next action. If confidence is "low" or backtrack_target_step_id is null, escalate to a human operator instead of proceeding autonomously.

This prompt is designed to be called by an agent orchestrator when the agent's current plan branch has exhausted all viable actions and no forward progress is possible. The orchestrator should invoke this prompt with the full plan history, the dead-end branch context, and the original goal. The prompt is not a standalone chat interaction; it is a structured recovery step inside an agent loop. The orchestrator must supply the agent's execution trace, the decision points encountered, and the specific branch that reached a dead end. Without this context, the model cannot identify valid backtrack targets.

Wire the prompt into your agent harness as a recovery handler. When the agent's action loop returns a terminal failure or an empty action set for the current branch, the harness should: (1) capture the full plan tree with completed, failed, and pending nodes; (2) serialize the dead-end branch including all attempted actions and their outcomes; (3) populate the [PLAN_TREE], [DEAD_END_BRANCH], and [ORIGINAL_GOAL] placeholders; (4) call the model with a low temperature (0.0–0.2) to maximize deterministic backtracking; (5) parse the output against the expected schema—a JSON object with backtrack_target_node_id, rationale, alternative_branch_suggestion, and lessons_learned; (6) validate that backtrack_target_node_id references a real node in the supplied plan tree and is not the dead-end node itself; (7) if validation fails, retry once with the validation error appended to [CONSTRAINTS]; (8) if the second attempt also fails, escalate to a human operator with the full trace and both failed outputs. Log every backtracking invocation, the selected target, and whether the alternative branch succeeded, so you can measure backtrack accuracy and dead-end recognition rates over time.

Model choice matters. Use a model with strong reasoning capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent) because backtracking requires causal analysis of why a branch failed and which earlier decision created the dead end. Smaller or faster models often suggest the most recent decision point rather than the true root cause, leading to repeated dead ends. If your agent runs with tool-use capabilities, do not expose tools during this recovery step—the prompt should produce a plan revision, not execute actions. After the backtrack target is selected, your orchestrator should reset the agent's state to that decision point, inject the alternative branch suggestion, and resume execution. Set a maximum backtrack depth (e.g., 3 rollbacks per task) to prevent infinite loops. If the agent exhausts all backtrack attempts, escalate the entire task to a human with the full plan tree and failure history.