Skip to content

Evaluate

Development notes for the upcoming Evaluate capability:

  • Evaluate agents shall belong to their own system-level namespace called _textevolve_evaluate_agents and need to be dynamically seeded prior to running an evaluation.
  • Agents that participate in the debate will not persist episodic memories of their own by design.
  • When building semantic memory collections for Evaluate, agents the corpus reader and re-voice process should be aligned with the agent’s general role in mind. For example, a Critic reading a corpus containing information about project management should capture facts from the corpus relevant broadly to the role of a critic. The re-voice process is used to fine-tune the corpus on the specific application of the information, for example, rewrite the memories through the perspective of a critic for large IT projects. Observe that there is a natural hierarchy of specialization emerging.