self-reflection

advancedwidespread

The AI agent reviews its own work, searches for errors, and makes improvements based on lessons learned before delivering a final result.

In plain terms

When you write an important email, ideally you do not send it right away: you read it through, notice a missing attachment or an ambiguous sentence, and correct it. Self-reflection builds this habit into the AI agent: the system does not stop at the first draft but looks back at it and asks itself whether it is truly good.

It is useful because the first response of language models is often incomplete or incorrect, yet the same model often notices the error if explicitly asked to critique it, especially if it receives tangible feedback, such as test results. This is particularly important for agents, which work in many steps, and an early mistake can steer the entire task in the wrong direction.

It differs from a simple 'try again' in that the agent puts into words what went wrong and why, and carries this lesson forward to the next attempt. It does not repeat blindly but works from its own critique.

How it works

The agent creates the first draft: it writes the text, the program code, or executes a step in the task.
It then switches to an evaluative role: it re-reads the result and compares it with the task, or uses external feedback, such as the error message from a test it has run or the results of a search.
It formulates criticism in everyday language, for example: 'I used the wrong date in the third step, so the total did not add up.'
It produces an improved version based on the critique and stores the lesson in its memory so it does not repeat the same mistake in the next round.
The cycle repeats until the result is satisfactory or until a set limit for attempts, time, or cost is reached.

An everyday example

You ask a coding agent to fix a bug on your webshop's cart page. It writes the fix, runs the tests, sees that one fails, determines it misread the rounding, corrects it, and only then tells you it is ready.

Who is affected

developerThe reliability of coding agents largely depends on whether they filter out their own errors through testing and self-checking before handing over the work. → Code

employeeIf you have an analysis or summary created with AI, a 'review and correct your own answer' loop often yields noticeably better results. → Work & HR

entrepreneur, executiveSelf-reflection improves quality but takes more time and money, so you must weigh which tasks are worth it.

parent, student, teacherAI self-checking is a good pattern for learning, but it does not replace your own verification of the received answer. → Education & Family

A common misconception

MythSelf-reflection means that the AI has become self-aware and thinks about itself like a human.

In realityThis is a working method, not consciousness: the system receives its own output as input and writes a critique about it. There is no question of self-awareness or feelings.

Limits and risks

The agent reviews its work with the same knowledge it used to do it, so if it missed something the first time, it is likely to overlook it again.
Without an external checkpoint (test, source, calculation, human feedback), the critique can easily become confident self-justification or even ruin a good answer.
Every reflection loop means more computation, so the response is slower and more expensive.
The agent can get stuck in an infinite loop of corrections, so a maximum limit for iterations is always required.

Similar, but not the same

reasoning model – A reasoning model thinks through steps before answering, while self-reflection reviews an already completed result afterward.
human oversight – In self-reflection, the agent checks itself; in human oversight, an independent person approves or stops the work.
fine-tuning – Fine-tuning permanently rewrites the model's internal settings, while self-reflection only corrects during the specific task using text notes, without changing the model itself.

Related concepts

Broader

AI-agent, agentikus munkafolyamat, Agentic design patterns

Narrower

Reflexion method, self-improving loop, evaluator–refiner pattern

On the edge

chain of thought, planning, agent memory, hallucination, human oversight

A brief history

The method was popularized by two 2023 studies: Reflexion (Shinn et al.) and Self-Refine (Madaan et al.) showed that language models achieve better results by learning from their own textual critiques. In 2024, Andrew Ng promoted it as one of the four fundamental agentic design patterns.

Where things stand

By late 2026, the self-check loop is a standard part of coding and research agents, mostly paired with automated tests or a separate checking agent. Developers are now mainly looking for ways to make the review reliable without raising costs excessively.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Nvidia's China share falls as it weighs Reflection AI deal and unveils agent security platformOct 5, 2026 · Agent Age
Hungarian AI Act guides: KPMG, SME checklist, lawyer interview, Blikk seriesAug 27, 2026 · Jog

Also known as

self-reflection · reflection · self-critique · self-check · self-reflection · self-critique · reflection · agent self-reflection

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.