agent-planners

Add structured planning to Python agents by injecting deterministic planning instructions and maintaining planning state across LLM calls using BasePlanner methods. This enforces predictable output structure and consistent reasoning before execution.

ReAct-style planning enforces agent output structure by requiring explicit /*PLANNING*/ and /*FINAL_ANSWER*/ blocks. This separation ensures clearer intermediate reasoning and safer final responses during LLM-driven task execution.

Maintain a task board for agent orchestration by storing task status in ctx.state and injecting it into the system prompt. This approach pairs with ManageTasksTool actions to drive tool-based task execution.

Yes, you can generate custom planning instructions per request by implementing BasePlanner. It allows you to inject optional planning instructions dynamically, adapting the agent reasoning structure to specific workflow needs.

Use task board planning when you need maintainable, tool-driven task status updates across steps, and choose ReAct-style planning blocks when you need explicit intermediate reasoning and safer final responses without external task management.

No external dependencies are required to manage task state. The Skill handles task state injection natively through ctx.state with ManageTasksTool actions, maintaining task status directly within the agent workflow.