11 Agent planning and feedback

This chapter advances the book’s treatment of reasoning by putting planning at the center of agentic systems. It distinguishes simple chat interfaces from true agents that decompose a user goal, decide on ordered steps, act through tools, and return results. The text shows why planning matters—especially for multi-step, dependency-laden tasks—and notes that some commercial platforms increasingly hide this complexity by performing sequential plans under the hood. The chapter’s scope spans fundamentals, hands-on demonstrations, construction of a general planner, and the role of feedback as a critical but often overlooked ingredient.

Concretely, it contrasts raw LLM tool use—good for parallel, independent actions—with agents that can execute sequential plans. Using Nexus and OpenAI Assistants as running examples, it illustrates how disabling planning strands an LLM after the first tool call, whereas enabling a planner achieves the end-to-end goal. The chapter then builds a generic, prompt-driven sequential planner that emits compact JSON task lists (including a for-each iterator), executes them locally, tracks outputs in a shared context, and finally asks the LLM to summarize results. Executing plans outside the model reduces token and tool overhead and works even when the LLM lacks native planning, though it introduces safety, tool-limit, and alignment considerations and requires selecting only the minimal actions needed.

Beyond planning, the chapter elevates feedback and evaluation as pillars alongside reasoning. It reviews OpenAI’s Strawberry (o1-preview) as a step toward models that integrate planning, reasoning, and self-critique, while showing that even such models can err and still benefit from guided feedback prompts. Practical guidance maps where to locate planning, reasoning, evaluation, and feedback across applications—personal assistants, customer service, autonomous agents, collaborative workflows, game AI, and research—balancing latency against depth of thought and deciding what to keep inside the LLM versus externally monitored. The chapter closes with exercises on building planners, adding feedback loops and error handling, and comparing parallel versus sequential execution, and it looks ahead to tighter model-level integration of these components.

The agent planning process

Creating a new agent in Nexus

Selecting the actions for the agent and disabling the planner

The results from trying to get the agent/LLM to complete the goal

The difference between iterative and planned execution

The assistant processing the goal and outputting the results

The planning process for creating and executing a plan

The results from requesting to complete the goal in Nexus using the basic planner

The time travel problem, revisited

• Planning is central to agents and assistants, allowing them to take a goal, break it into steps, and execute it. Without planning, agents are reduced to simple chatbot-like interactions.
• Agents must differentiate between parallel and sequential actions. Many LLMs can handle parallel actions, but only advanced models support sequential planning, critical for complex task completion.
• Feedback is crucial in guiding agents to correct their course and improve performance over time. This chapter demonstrates how feedback mechanisms can be integrated with agents to refine their decision-making processes.
• Platforms like OpenAI Assistants and Anthropic Claude support internal planning and can execute complex, multi-step tasks. Agents using these platforms can leverage sequential action planning for sophisticated workflows.
• The proper selection and limitation of agent actions are vital to avoid confusion and unintended behavior. Too many actions may overwhelm an agent, while unnecessary tools may be misused.
• Nexus allows for creating and managing agents through a flexible interface, where users can implement custom planners, set goals, and assign tools. The chapter includes practical examples using Nexus to highlight the difference between a raw LLM and a planner-enhanced agent.
• Writing custom planners is straightforward, utilizing prompt engineering strategies. Tools like LangChain and Semantic Kernel offer a variety of planners that can be adapted or extended to fit specific agentic needs.
• Models like OpenAI Strawberry integrate reasoning, planning, evaluation, and feedback directly into the LLM, offering more accurate problem-solving capabilities.
• Evaluation helps determine how well an agentic system is performing and can be implemented internally or externally, depending on the use case.
• As LLMs evolve, reasoning, planning, and feedback mechanisms may become deeply integrated into models, paving the way for more autonomous and intelligent agent systems.

FAQ