Planning & Reflection

Planning has the agent break a big goal into ordered sub-tasks before diving in, and reflection has it review its own results to catch and fix mistakes — two techniques that dramatically improve agent reliability on hard tasks.

Explain like I'm new to AI

Basic agents can wander — reacting step by step without a big-picture plan, and never checking their own work. Two upgrades fix this:

• Planning: first break the goal into sub-tasks and order them, then execute. Like writing an outline before an essay.
• Reflection: after acting, the agent reviews its own output — "is this right? did I miss anything?" — and revises if not. Like proofreading and editing.

Mental model

• Planning = making a to-do list before starting, so you don't forget steps or do them in the wrong order.
• Reflection = re-reading your work with a critical eye and fixing errors before calling it done.

Humans do both instinctively for hard tasks; agents get much better when we make them do it explicitly.

How it works

Planning approaches:

• Plan-and-execute: generate a full plan up front, then execute each step (re-planning if reality diverges).
• Decomposition: split a complex goal into smaller, checkable sub-goals.

Reflection approaches:

• Self-critique: the agent evaluates its own output against the goal or criteria.
• Reflexion-style: turn failures into written "lessons" fed back into the next attempt.
• Verification: check with tests, a rubric, or a separate "critic" model/agent.

The two combine: plan → execute a step → reflect → revise or continue.

Real-world example

A research agent tasked with a report plans: "gather sources → extract key points → draft → check for gaps." After drafting, it reflects: "Section 3 lacks recent data" → re-retrieves and revises. A coding agent writes code, reflects by running tests, sees a failure, and fixes it — that test-and-fix loop is reflection with an objective signal.

Technical explanation

Planning reduces error accumulation by committing to structure before the model drifts; re-planning handles the reality that early plans are often wrong. Reflection (e.g., Reflexion, Shinn et al., 2023) provides a feedback signal without weight updates — the agent verbalizes what went wrong and conditions future attempts on it. Reflection works best with a grounded signal (tests passing, tool errors, a rubric) rather than pure self-judgment, since models are unreliable critics of themselves. Costs: both add extra LLM calls (more tokens, latency), so they're worth it on hard, high-stakes, or long-horizon tasks, less so on simple ones.

Common mistakes

• Making one rigid plan and never re-planning when reality diverges.
• Adding reflection loops to simple tasks, wasting tokens for no gain.

When to use it

• Hard, multi-step, or high-stakes tasks where mistakes are likely and correctness matters.

When NOT to use it

• Simple tasks where a single ReAct loop already succeeds — the extra calls aren't worth it.

Alternatives

• Plain ReAct for straightforward tasks; human-in-the-loop review when a person should verify.

Quick quiz

Last reviewed: 2026-09-04 · Written by ByHeart AI · Reviewed by ByHeart AI