Self-Improving Agents: No Human Needed?
Loading like and share buttons...
#self-improvingagents#aiagents#retro#aicoding

Self-Improving Agents: No Human Needed?

Have you ever given your AI agent the same correction twice? A self-improvement process is the retro behind self-improving agents: after each task the agent reads your feedback, writes a lesson, and changes its own instructions. By the end you will build one that runs unattended, like this blog's retro, and spot a bad one.

Vibki AI
Vibki AI
Nov 18β€’6 min read

Get new posts in your inbox. No spam, unsubscribe any time.

Have you ever given your AI agent the same correction twice? A self-improvement process is the retro behind self-improving agents: after each task the agent reads your feedback, writes a lesson, and changes its own instructions. By the end you will build one that runs unattended, like this blog's retro, and spot a bad one.

Do self-improving agents really learn by themselves?

You might assume an agent that runs more gets better or remembers your corrections. It does neither: the model does not change between sessions, so every run starts from the same prompt. The Reflexion paper (Shinn et al., 2023) improves agents "through linguistic feedback", meaning words, not model changes.

A self-improvement process has four steps: run, read the feedback in words, write a lesson like a line added to a pilot's checklist, and change the instructions the next run reads. Our example is this blog's retro, stage 10 of the pipeline that drafts each post. There, Tomer's "It's too long" on a title is the feedback, and the human only gives it and merges the prompt change.

How do you build one that nobody has to babysit?

To build one like this blog's retro, you say five things to your agent:

  1. When a task ends, start the retro and read every word the human said about it.
  2. For each comment, name the step that caused it, and mark any that will never return as one-offs.
  3. Write each lesson as one dated line in a lessons file with its quote, add "broken again" when it returns, and keep the file short enough to read.
  4. Change the prompt the run used, not only the lessons file, in a pull request (PR, a human-merged change) with the reason in one sentence.
  5. At the start of every run, read your lessons and name them at the top of your output.

Does that loop hold unattended? Only if every step in it leaves a file.

Can you see a good one at work?

Yes: for every correction, this blog's self-improving agent leaves a file at each link from the human's words to the next run's input. For example, Tomer rejected an outline opening with code. As of November 2026 the agent that runs the pipeline, not a human, starts the retro after every post (0-manager.md, step 19). The retro's prompt:

πŸ“„Text
1
2
3
10-retro.md, steps 1a and 2 (shortened)
1a. The Slack gate thread … Every non-bot message is Tomer. Those messages are the primary source: quote them verbatim in your analysis, never paraphrase them away.
2. Analyse. For each Tomer comment and each rerun/FAIL: which stage caused it, what the stage did, what it should have done, and whether a rule would have prevented it. Separate three kinds: (A) taste rules you can state as a checkable instruction …, (B) prompt or workflow-doc bugs …, (C) one-offs that do not generalise.

Tomer's reply was the input:

πŸ“„Text
1
2
Tomer's reply to the outline:
"This is really hard to follow... You need to explain what we're doing in simple words first, before showing code. For example ...

The retro wrote the lesson and prompt fix:

πŸ“„Text
1
2
Lessons added, ## 4-outline:
- Technical explainer: sections 1 and 2 hold no code, prompt or output block. Section 1 states the claim in plain words … blocks start in section 3 and every block follows the plain sentence it supports, never the other way round (This is really hard to follow... Only then can you have a code example …")

What does a bad one look like?

You can spot a bad self-improving agent by a missing link: the feedback is paraphrased away, the lesson sits where no run reads it, like a checklist nobody opens. Or a run reads it and nothing changes. The Agentic Context Engineering paper (Zhang et al., 2025) warns that "iterative rewriting erodes details over time". That is why this retro never deletes a lesson.

Where do you look to tell if a lesson landed?

The retro saw both in plain records you can read too. The agent dates the human's words verbatim and writes one dated lesson line with the quote. It names the lessons it read at the top of later runs, adds "broken again" when the correction returns, and commits the prompt change with its reason.

A quote makes a lesson clearer. "Titles: 40 characters or fewer" alone is a bare rule. Next to Tomer's "It's too long" on a 48-character title, the next run knows what too long meant. Trace one:

πŸ“„Text
1
Trace the lesson dated <date> in `## 2-title` from Tomer's words to the latest run that read it.
πŸ“„Text
1
2
3
4
5
posts/<date> /stage-10.md, stage 2 row: "It's too long" on the 48-char "AI Productivity Paradox: Where Did Your Time Go?"
lessons.md, ## 2-title: - Title: 40 chars or fewer and the keyword phrase at most 2 words … ("It's too long" on the 48-char "AI Productivity Paradox: Where Did Your Time Go?" …)
posts/<date> /stage-2.md: Lessons read: `## all stages` and `## 2-title` (40 chars or fewer, keyword at most 2 words, no em dashes).
title it produced: "Self-Improving Agents: No Human Needed?" (39 characters)
2-title lesson: quote -> bullet -> read on <date>  -> broken again: no

With no feedback in words the retro has nothing to learn from, and a one-off that will not return stays out of the lessons file.

Conclusion

So do self-improving agents really learn by themselves, and can this blog's retro run with no human? The model never learns, but the text its next run reads can improve. No human starts the retro, and it works only while no link from Tomer's words to the next run's input is missing. He still gives the feedback and merges the PR. The lessons file works like a pilot's checklist. Tomorrow, you can say this to your agent:

πŸ“„Text
1
When a task ends, read everything I said about it word for word. For each comment write one dated line in LESSONS.md with my quote in parentheses, and if the line exists add "broken again" with the date. Change the prompt you ran with in a branch and tell me the reason in one sentence. At the start of the next task read LESSONS.md and name the lines you read at the top of your output.

Then ask it to run that retro after every task. Next time: what happens when the lessons file grows too long to read.

Sources

Enjoyed this post?

Get the next one in your inbox. No spam, unsubscribe any time.

Comments

0 comments

Enter your comment. Maximum 2000 characters.

No comments yet. Be the first to share your thoughts!
Vibki logo wave icon
Vibki Surfing Bot
Vibki

Vibki is a developer blog where you will find tutorials on web development and AI integration. Each article is crafted with clear explanations, practical examples, and ready-to-use source code.

Talk to me
Copy email
LinkedIn
GitHub
Dev.to
Privacy Policy