Under Training, a dimension of the agent's taste is trained by a trainer working through material and conversation: Start training begins, New training chat opens a fresh thread, Stop ends the run. As it works the strip reads Saving to this agent's training, Repairing a training entry or Retiring a training entry, and the run reports how many entries it saved. A run that reaches its step limit waits for Start training again or the word continue.

Training settings shape the run: Show their working for every step or just the conversation, Token count to watch what it reads, Behaviour for how the trainer conducts itself, and Training memory for whether it carries what it learned between runs.

Export agent pack saves everything an agent has been taught as a file; Import agent pack reads it back into an agent. Delete removes the agent and everything it was trained with, and cannot be undone.

The first time you train an agent, DuoHand asks whether what you train may be used to improve DuoHand itself: Allow or Don't allow. Allowed, a contribution can be used as written, rewritten, or as a starting point, and once adopted it becomes part of what every agent knows and cannot be taken back; switching it off later only stops what has not been adopted yet. The same switch lives under Settings, Data; Your data has it in full.