Iterative Planning and Re-planning with Prompt Adjustments

AI agents operating in dynamic environments often require their initial plans, no matter how carefully constructed, to be modified. Unforeseen obstacles, new information, or failures in tool execution can render parts of a plan obsolete or suboptimal. Iterative planning and re-planning are essential for managing these challenges. Crafting prompts that guide an agent through these adjustments is fundamental to building resilient and effective agentic systems. Prompt engineering techniques that enable agents to adapt their plans in response to changing circumstances are examined.

The Need for Adaptive Planning

"Agents executing tasks in or complex digital environments rarely have the luxury of a perfectly predictable path. A web service might become unavailable, a data source might return unexpected results, or a user might introduce a new constraint mid-task. Without the ability to re-plan, an agent would either fail or produce an irrelevant outcome. Iterative planning allows an agent to:"

• React to failures: If a planned step fails (e.g., a tool error), the agent needs to find an alternative.
• Incorporate new information: New data might reveal a more efficient path or invalidate a current assumption.
• Optimize dynamically: As the task progresses, opportunities for improvement or shortcuts might emerge.

Prompts are the primary mechanism through which you, the developer, can instruct an agent on how to approach these situations.

Triggering and Guiding Re-planning with Prompts

When an agent encounters a situation that requires a deviation from its current plan, a specific type of prompt, often called a re-planning prompt, is used. This prompt's purpose is to provide the agent with the necessary context and instructions to generate a revised plan.

Main components of an effective re-planning prompt include:

1. The Original Goal: Reinforce the overarching objective to ensure the re-planning effort remains aligned.
2. The Current Plan State: Detail the original plan, which steps have been completed, and which step encountered the issue.
3. The Problem Description: Clearly state the error, new information, or obstacle that triggered the re-planning. This might be an error message from a tool, feedback from a sensor, or new input from a user.
4. Relevant Contextual History: Include any recent observations or short-term memory items that might influence the re-planning decision.
5. Specific Instructions for Re-planning: Guide the LLM on how to approach the re-planning task. This could involve asking it to:
   • Analyze the cause of the failure.
   • Propose alternative actions or sub-goals.
   • Examine specific constraints or preferences for the revised plan.
   • Evaluate trade-offs if multiple solutions exist.

Here's an example structure for a re-planning prompt template. Imagine this is part of a Python script that dynamically fills in the details:

# Re-planning Prompt Template Example
# This would be populated with actual data at runtime.

original_goal = "Book a flight from New York to London for next Monday, economy class, budget $500."
previous_plan_steps = [
    "1. Search for flights on AirlineX (Completed)",
    "2. Identify suitable flights under $500 (Completed)",
    "3. Attempt to book flight AX123 (Failed)"
]
current_step_action = "Attempt to book flight AX123"
error_message = "AirlineX API returned: 'Seat not available at selected price'."
new_information = "User indicated a willingness to book flights up to $550 if direct."

re_planning_prompt = f"""
System: Agent Re-planning Module

Original Goal: {original_goal}

Previous Plan Execution:
{chr(10).join(previous_plan_steps)}

Current Step Attempted: {current_step_action}
Outcome/Error: {error_message}
New Relevant Information: {new_information}

Task:
The current plan to book flight AX123 has failed due to unavailability at the desired price.
1. Analyze the situation.
2. Propose a revised plan to achieve the original goal of booking a flight.
3. Your revised plan should incorporate the new information about budget flexibility for direct flights.
4. If multiple options arise, prioritize direct flights if they fit the adjusted budget. Otherwise, find the cheapest option within the original $500 budget, possibly on a different airline or with a layover.
5. List the steps of your revised plan clearly.

Revised Plan:
"""

This prompt doesn't just state the problem; it actively guides the agent's thought process for generating a solution. The "Prompt Adjustments" here refer to constructing these tailored prompts on-the-fly, adapting a base template with specific, current information to address the immediate re-planning need.

Strategies for Iterative Planning Prompts

Several strategies can make your re-planning prompts more effective:

• Scoped Re-planning: Instead of asking the agent to re-plan everything from scratch, you can instruct it to focus on modifying only the necessary parts of the plan. This can save computational resources and maintain stability in the parts of the plan that are still valid.

  • Example: "Revise only the next 1-2 steps to overcome this obstacle. The subsequent parts of the plan should remain unchanged if possible."

• Explicitly Requesting Justification: Asking the agent to explain its reasoning for the revised plan can help in debugging and also encourages more logical plan generation.

  • Example: "For each step in your revised plan, briefly explain why it was chosen and how it addresses the issue."

• Providing Fallback Options or Hints: You can embed suggestions or preferred fallback strategies directly into the prompt if certain failures are common or have known solutions.

  • Example: "If searching on DatabaseA fails, try DatabaseB as an alternative before reporting a failure to find the information."

• Managing Re-planning Depth: For complex problems, an agent might get stuck in deep re-planning loops. You can introduce limits or criteria for escalating the issue if re-planning fails after a certain number of attempts.

  • Example: "If this is the third attempt to re-plan for this specific sub-task and a viable solution is not found, report the impasse and list the unresolved constraints."

The following diagram illustrates the flow of an agent's execution cycle, highlighting where iterative planning and re-planning fit in.

An agent's execution cycle. If a step's outcome (4) is not as expected, a re-planning prompt (5a) is constructed. This prompt, containing the goal, current plan, and the issue, guides the LLM to adjust the plan (5b), which then feeds back into the execution loop.

Approaches for Adaptive Planning

When designing prompts for iterative planning and re-planning, keep these points in mind:

• Clarity of Failure/Change: The agent needs unambiguous information about what went wrong or what has changed. Vague error reporting in your prompts will lead to vague or incorrect re-planning.
• Balancing Autonomy and Guidance: Prompts should provide enough guidance to ensure the re-planning is productive but also allow the agent some flexibility to find creative solutions. Overly prescriptive prompts can stifle the LLM's problem-solving capabilities.
• Resource Consumption: Re-planning, especially if it involves multiple calls to an LLM, consumes tokens and time. Prompts can be designed to encourage efficient re-planning, for example, by asking for minimal necessary changes.
• State Management: Ensure that the agent's internal state (what it knows, what it has tried) is accurately reflected in or accessible to the re-planning process. This is important for avoiding repetitive failures or cycles.

By mastering prompts for iterative planning and re-planning, you enable your AI agents to navigate the complexities and uncertainties of their tasks more effectively. This adaptability is a hallmark of more sophisticated agentic systems, allowing them to recover from setbacks and dynamically adjust their strategies to achieve their objectives. The next sections will build upon these planning capabilities, exploring how agents evaluate the quality of their plans and manage overall task execution.