Continuous Agent Evolution

Paradox: an agent solves new tasks without examples, but after ten thousand similar ones can repeat the mistake of the first day. Preserving experience ≠ learning from experience: recording trajectories in a database does not mean extracting patterns from them. Learning happens after “evaluation, matching, generalization, and verification,” not at the moment of writing to disk. As long as the model is not capable of reliably learning continuously, learning is formalized as an external system: record evidence, verify, extract patterns, choose an update carrier (knowledge/instructions/programs/parameters), and run candidate versions through regression.

Related: [Learning signals from trajectories], [Four update carriers], [Two evolution loops]