Learning from deployment
reinforcing loop · 4 steps · Trace on the map →
The cycle
- Agents do real workLong-horizon agents
- Traces of what workedTrace capture
- Agents learn from themContinual-learning agents
- Better agents get more work
What drives it
Work produces the experience that makes agents better at the work.
What breaks it
It needs a trustworthy signal of what "worked", so it stalls where verification is weak. Today most of the learning lives outside the weights, in files and memory.
What to watch
Whether labs ship weight-level learning from deployment. How fast an agent improves inside one company.