The loop most agents skip: learning from outcomes.

Most AI agents run. They take an instruction, act, and stop. A smaller number remember what happened last time. A far smaller number change what they do because of how the last attempt turned out. The distance between an agent that runs and an agent that improves is not model size or prompt length. It is whether the system can tell, after the fact, whether the work was any good, and whether that verdict is allowed to change the next run.

That verdict is the part teams skip. Capturing an outcome is easy to describe and easy to omit, because nothing breaks when you leave it out. The agent still runs tomorrow. It runs the same.

Measurement is the whole game

"Self-improving" is a claim that only means something when the system has a way to know it improved. Without a score, accumulation is not learning. More stored context, more remembered edits, more saved prompts: each is as likely to add noise as signal, and noise compounds quietly. The public writing on recursive improvement keeps returning to one requirement, an automated evaluator that can rank a candidate change before it goes live. The mechanism differs by source; the precondition does not.

So the useful loop is narrow:

• The agent acts.
• A human corrects, approves, or rejects, and where possible the result is measured (a reply came back, a draft went out with few edits, a check passed).
• That outcome becomes one durable thing: a memory, a rule, or a reusable skill.
• The change is promoted only if an evaluation says it beat what it replaced.
• A human approves anything that leaves the building.

Step 4 is the one that separates a system that sharpens from one that drifts.

What the loop captures

The raw material is the outcome, not the output. An output is the draft the agent produced. An outcome is what happened to that draft: the edit distance between what the agent wrote and what the person sent, whether a proposal was accepted, whether a follow-up got a reply. Outcomes are harder to capture because they arrive later and out of band. They are also the only signal that carries information about quality rather than volume.

A durable learning is a short, testable statement. "For this reader, lead with the number before the method." "This client wants milestone pricing, not hourly." Each one is versioned, each one can be rolled back, and each one is attached to the outcome that justified it. That audit trail is what makes the loop safe to run without a person watching every step.

Rarefied Earth as client zero

We run pieces of this on our own operation before we would offer any of it. A candidate queue records reusable work as it happens. A verification ledger tracks what has been checked. A voice guard is a hard gate that refuses prose containing the patterns we have banned, and it cannot be bypassed by asking nicely. A weekly loop researches one real work item and harvests it three ways.

Here is the honest limit. We have not packaged "an agent that learns from its own outcomes" as a product a client installs. The bar for that is a named pain and a system that has run in daily operation long enough to be trusted, not a good idea written down. This piece describes the pattern and how we hold ourselves to it. It does not describe a shipping feature you can buy today. When it becomes one, it will have earned the label by running here first.

What you can do with one workflow

Pick one repeated task where the output is text and the outcome is observable: outbound follow-ups, proposal sections, meeting prep. Define, before you automate anything, the single measurement that means "better" for that task. Capture the outcome every time. Let corrections become versioned rules, and promote a rule only when the measurement moves in the right direction. Keep a human approval on anything that reaches a customer. Extend to a second workflow only after the first one shows the number improving.

The goal is not autonomy. The goal is a system that compounds what it learns instead of forgetting it, with a gate honest enough to reject its own bad ideas.