How should an agent plan and replan during a task?

Uses plans as bounded, revisable execution aids while preserving evidence, policy, and application-owned state transitions.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

For a simple task I let the agent choose one next action at a time. For a longer task, a short explicit plan can expose dependencies, required evidence, and completion criteria. The application validates eligible actions and updates durable task state after each result. Replanning happens when evidence invalidates the current path, not after every minor observation.

Full answer below

Practice focus

Planning strategy, adaptation, progress, evidence, and control of model-generated plans.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
A model-generated plan is a proposal for execution, not an authorization or durable source of truth.
For a simple task I let the agent choose one next action at a time. For a longer task, a short explicit plan can expose dependencies, required evidence, and completion criteria. The application validates eligible actions and updates durable task state after each result. Replanning happens when evidence invalidates the current path, not after every minor observation.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
Plans use stable step identifiers and distinguish pending, running, completed, skipped, blocked, and failed work. Completed side effects are never erased by a rewritten plan. The agent explains a plan change in terms of new evidence or a failed prerequisite, while the runner checks remaining budget and policy. Independent steps can run in parallel only when their tools and outputs do not conflict.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Treating the initial plan as immutable.
Erasing completed side effects when replanning.
Persisting private model reasoning as the audit record.

How should an agent plan and replan during a task?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue practice

Next in path

Tech stacks

Skill areas

Human editorial direction

Was this drill useful?

Found this answer useful?