A writer who rereads a draft and fixes it is self-correcting. For a model, the simplest version is a second request: “Here is your answer. Check it against these requirements and fix any problems.” Self-Refine studies exactly this loop, with one model drafting, critiquing and rewriting (Madaan et al.). Reflexion, another research method, has an agent write a short reflection after a failed attempt and keep it in memory for the next try (Shinn et al.). “Reflection” is the common informal name for both ideas.
The catch is where the feedback comes from. Huang and colleagues found that when models judged their own reasoning with no outside signal, they often failed to fix errors and sometimes changed right answers to wrong ones (Huang et al.). Self-correction is far more dependable when something external says what went wrong: a failing unit test, a compiler error, a schema validator, a retrieved source or a separate LLM-as-a-judge with clear criteria. In product terms, build it as a prompt chain with logged steps, which is how Anthropic’s prompting guide describes its most common chaining pattern (Anthropic, Prompting best practices).
In episode 52 at 39:10, Airbyte’s Michel Tricot argues that it is fine for a coding agent to make bugs: “We’re just looking for something that can create massive throughput and can self-correct.” That is his view of where agent adoption is heading, and it assumes an environment, such as tests and builds, that tells the agent when it is wrong. Galileo was the presenting sponsor of episode 52 (opening read).
Self-correction does not remove the need for checking. A revised answer can still contain an AI hallucination, so measure the final output, not just whether a revision happened.
Hear it from the guest
“We're just looking for something that can create massive throughput and can self-correct.”
Quotes lightly edited to remove filler words.
Sources
- Madaan et al.: Self-Refine — Uses a single LLM as generator, feedback provider and refiner, evaluated across seven tasks without extra training; checked October 24, 2026.
- Shinn et al.: Reflexion — Agents reflect verbally on task feedback and keep that reflective text in an episodic memory buffer for later attempts; checked October 24, 2026.
- Huang et al.: Large Language Models Cannot Self-Correct Reasoning Yet — Finds that LLMs struggle to self-correct reasoning without external feedback, and that performance sometimes degrades after self-correction; checked October 24, 2026.
- Anthropic: Prompting best practices — Describes the most common chaining pattern as self-correction: draft, review against criteria, refine, with each step a separate API call; checked October 24, 2026.
- Chain of Thought episode 52 at 39:10 — Michel Tricot says coding agents should be judged on throughput and their ability to self-correct, not on never making bugs.
