Guide
Do AI agents learn from experience?
Yes — and almost never in the way the question implies. The model does not get smarter. The system around it gets better informed. Once you see where the line falls, a lot of confusing vendor language resolves, and you can predict what an agent will and will not pick up.
The short answer
An agent learns from experience, but its model does not. The weights inside the language model are fixed; they were set before you ever used it and nothing you do in normal operation changes them. What learns is the material the model reads before it answers.
That sounds like a technicality. It is the single most useful thing to understand about agents, because it tells you precisely what will improve and what will not.
The four things that actually change
- Memory. Facts about your business, your preferences, decisions already made. The agent that knows your warranty is twelve months knows it because it was stored, not because it worked it out again.
- Skills. A procedure captured once and replayed identically. This is where most real improvement lives — see how agents learn new skills.
- Knowledge. The documents it can consult. Adding a price list is learning, in every sense that matters to you.
- Instructions. Its own operating text, refined over time as you correct it.
None of that is fine-tuning, and conflating the two leads to bad decisions in both directions — memory vs fine-tuning.
What this predicts
The framing earns its keep here, because it lets you forecast behaviour instead of hoping.
An agent will reliably get better at: knowing facts about your business, following a procedure you have specified, using your vocabulary, and remembering decisions and preferences. All of these are storage and retrieval problems, and storage works.
An agent will not get better at: reasoning that was beyond the model on day one, judgement in genuinely novel situations, or anything requiring a capability the model does not have. No amount of experience teaches a model a skill it lacks — it can only be told, in more detail, what to do.
So when an agent keeps getting something wrong, the diagnostic question is: is this a knowledge problem or a capability problem? Knowledge problems are fixable by writing something down. Capability problems need a different model, or a human.
Why this is good news
Three consequences follow, and all of them favour the buyer:
- Improvement is inspectable. The changes are text and files, so you can read what the agent learned. A weight update would be unreadable by anyone.
- Improvement is reversible. A wrong procedure can be deleted. A wrongly fine-tuned model has to be retrained.
- Improvement is portable in principle. A written procedure is not welded to one model — though whether a given platform lets you take it is a separate question: are agent skills portable.
How the learning actually happens
Two routes, and a good system has both. You teach it — correct it, write down the procedure, upload the document. Direct, immediate, and how most useful learning starts. It distils — reviews work it has already done and writes down what recurred, on a schedule rather than on demand.
The second is where "self-improving" comes from, and it needs bounding for the reasons set out in the pillar. The first needs nothing but you noticing a correction you have made twice.
A practical habit
The highest-return thing you can do with an agent takes about a minute: the second time you correct the same thing, write it down as a procedure instead of correcting it again.
That is the whole discipline. Teams that do it end up with agents that feel like they know the business; teams that do not end up re-explaining forever and concluding the technology does not work.
Related reading
Self-improving AI agents
The pillar: what learns, what can go wrong, and the safeguards.
Memory vs fine-tuning
Two different things that get called learning.
Teach an agent a skill
Turning a correction you keep repeating into a procedure.
How agents remember
The memory architecture: conversation, long-term, knowledge and skills.
FAQ
Do AI agents learn from experience?
Yes, but the model's weights never change. What learns is the material around the model — accumulated memory, written procedures and skills, the documents it can consult, and its own instructions. The model is the same one it was yesterday; the system it sits in knows more.
Does an AI agent get smarter over time?
Better informed, not smarter. It will reliably improve at knowing facts about your business, following procedures you specify, using your vocabulary and remembering decisions — all storage problems. It will not get better at reasoning that was beyond the model on day one, because experience cannot teach a model a capability it lacks.
Is agent learning the same as fine-tuning?
No. Fine-tuning changes model weights and is expensive, slow and opaque. Agent learning changes memory, skills and instructions — cheap, fast, readable and reversible. Conflating the two leads to bad decisions in both directions.
How do I teach an AI agent something?
Two routes. You teach it directly — correct it, write the procedure down, upload the document. Or it distils, reviewing work it has already done and recording what recurred. The most useful habit is simple: the second time you correct the same thing, write it down as a procedure instead of correcting it again.
An agent that keeps what it learns
Persistent memory, written procedures and skills distilled from real work — inspectable, reversible, and yours.