Online planning dynamically generates and revises an action plan while an agent executes a task. It matters when a workflow contains multiple tools, decisions, and possible failures rather than one predictable operation. section 9.2.2 section 9.3.1
An agent can activate planning through a prompt that asks a language model to reason with a specified method. Planning may decompose a goal into smaller subtasks, reflect on previous results, and store conclusions in memory. In production, explicitly coded workflows often replace probabilistic online planning because they reduce error potential. section 9.2.2 section 9.3.1
A practical planning loop can include:
- Decompose the goal into manageable subtasks instead of attempting the entire workflow at once. section 9.2.2
- Choose a sequence of tool calls that gathers relevant information and produces the intended result. section 9.2.2
- Reflect on action results and use the conclusions to revise behavior or update memory. section 9.2.2
- Add explicit controls where long-horizon reasoning, unexpected errors, or tool-call mistakes could break the workflow. section 9.3.1
Product roadmap update
A road-mapping agent first fetches the existing roadmap, searches for relevant trends and competitor activity, gathers and analyzes user feedback, and then uses a custom analysis tool to align suggestions with company priorities. Reflection can evaluate feedback from an individual subtask before the agent continues. section 9.2.2
