Incremental Plan Updates

Update the execution plan incrementally rather than re-planning from scratch on every turn. This avoids re-consuming the full context and keeps the agent focused on the remaining work.

How It Works

After the initial plan is created, each turn modifies only the parts that have changed: marking steps complete, adding new steps discovered during execution, or adjusting order based on new information. The plan is stored in state and injected into the prompt as a compact representation.

Benefits

• avoids re-sending the full problem description on every turn
• reduces the model's cognitive load by giving it a focused delta to reason about
• prevents the agent from oscillating between different strategies

Anti-patterns

• Re-planning from scratch after every tool call.
• Including the full original user request plus all prior reasoning in every turn instead of a concise plan state.
• Letting the model hallucinate new plans that contradict already-completed work.

See Also

• Stateful Tracking in Agent Loops — plan state is a key component of agent state.
• Compaction — the plan is part of what must be preserved during compaction.
• Structured State Extraction — the general technique this is a special case of.

Tools

• LangGraph — supports persistent state and incremental updates across agent loop steps.
• AutoGen — multi-agent framework where plans can be maintained and updated by a dedicated planner agent.