An agent can form the wrong lesson, then fix it after another correction. But what if it keeps making the same kind of mistake when forming new lessons? Updating a Skill can fix one task. The bigger change is improving how the agent turns the next experience into something worth learning.
This time, it remembered
The second article in this series asked why an agent might miss a saved lesson. This time, it finds the note and follows it. The trouble starts earlier: what lesson did it form from the experience?
Imagine asking for a quick update before a team meeting. The agent writes several pages. You say, “This report is too long.”
It saves: “The user prefers short reports.”
Later, you need a report to decide whether to spend money. The agent keeps it short, leaving out the evidence you need.
It remembered your words. But did it understand what they meant?
One comment can mean different things
Maybe you prefer short reports. Or maybe that meeting only needed a quick update. Both explanations make sense. One comment is not enough to settle it.
A belief is what we currently think is true. It shapes what we expect and do. If the agent believes you dislike detail, it will keep cutting detail. If it thinks the problem was specific to that meeting, it should check what the next report is for.
The event is evidence. The lesson is an interpretation of that evidence.
Now you ask for more evidence. Should it flip to “the user likes long reports”? That would repeat the same mistake with a different answer.
From experience to a Skill
There are three useful stages to picture here:
- Keep the event and its setting. What happened, and what were we trying to do?
- Find a possible pattern. Perhaps the amount of detail depends on the report's purpose.
- Turn it into a method. Check the purpose before deciding what to include.
Each step involves judgment. A shorter record can lose the setting. A pattern can stretch beyond the evidence. A method can turn a useful idea into an overly broad rule.
Agents already do parts of this. Hermes, an open-source agent, can create and edit Skills: reusable instructions, sometimes with scripts and supporting files. EvoMap's GEP protocol also describes strategies with conditions, checks, and outcome records that can guide later choices.
So the question goes beyond whether an agent can write or revise a useful procedure. It can. The question is what the revision changes.
Fix this lesson—or how the next one is formed?
The report-writing Skill could now say:
Keep meeting updates brief. Include enough evidence for spending decisions.
That fixes the lesson. But next week, you ask for one email to sound less formal. Will it save “the user always prefers casual writing”?
If so, the report rule improved, but the same mistake survived: turning one request into a lasting preference.
There are two places the result can feed back into:
- The lesson itself: change the rule about report length.
- The way lessons are formed: reconsider how the system separates a one-time request from a lasting preference.
We can add an instruction: “Do not save a permanent preference from one request.” That can help. But the system still has to decide what counts as enough evidence and when an exception matters.
If later results help revise that decision process, the feedback reaches further. If they only rewrite individual Skills, the system can keep repairing lessons while repeating the same mistake in how it creates them.
The agent updates the lesson. What updates how it forms the next one?
What people can learn beyond the rule
After a few misunderstandings like this, a person may stop taking every comment as a general rule. They start asking what the other person needed in that situation.
That change can involve four abilities:
- Notice the setting. A quick update and a spending decision need different evidence.
- Limit the conclusion. One request does not tell you what someone always wants.
- Revise a belief. New evidence can refine an explanation without simply reversing it.
- Carry the lesson across. Use that care with emails and presentations too, while noticing what is different.
Human studies show that practice can improve which differences people notice. In movement tasks, past errors can change how strongly people learn from later errors. These are specific examples of experience changing how later information is handled—not just what is remembered.
A growing library is only part of the story
An agent can improve its learning process through changes to its instructions, how it selects lessons, or how it uses feedback. The important question is where those changes reach.
If feedback stops at the saved lesson, improvement stops at that layer. The way the next lesson is formed does not improve automatically.
That is the gap a growing Skill library can hide.
Learning from experience should also improve how the next experience is understood. What would an agent need to keep making that change?
The final article asks what kind of system could keep making that wider change over time—and how much of it our work actually needs.
