The self-learning agent that everyone's talking about is not the one your product needs.
The most useful signal of all is actually one that almost nobody captures: the people using your product.
Your agents and users already work side by side, thousands of times a day.
It can learn from each one of these interactions. See the video.
VIDEO
This is a guide to all three learning layers & how it’s relevant to you.
Today, agents can learn in three different layers.
For each one: what it is, what it costs, who is actually building it, and whether you even should use self-learning.
You'll see how Anthropic, Karpathy, DeepMind, Microsoft, Hermes, OpenClaw, CopilotKit and others approach self-learning, and where each one stops.
---
## Agents can learn in 3 different layers
The cleanest split comes from Harrison Chase.
• The model = The weights (training parameters)
• The harness = The code around the model (loop, tools, prompts)
• The context = The memory and skills outside the harness which grow
You already use all three in Claude Code.
The model is Claude, the harness is Claude Code itself, and the context is your CLAUDE.md and your skills. Each layer can get better on its own, without touching the other two.
In a real product, self-learning almost always means the harness or the context, not the model. The model belongs to the labs. The other two are yours.
I will use one example across all three: a support agent that issues refunds.
Generated by Thread Navigator
Press ⌘ + S to quick-export