DigitalOcean has introduced services that turn deployed-agent interactions into material for evaluation and improvement. Learning Services records sessions, identifies performance gaps and runs simulated conversations so developers can test proposed fixes before exposing more users to them.
The product has three components. Insights captures agent activity, Signals surfaces failures, and Simulations exercises revised behavior against synthetic interactions. An enhanced inference router can compare model versions and choose among them based on task requirements, cost and quality.
Agent Data Services provides the supporting catalog for documents, audio, video, logs and session records, with search, governance and access controls. DigitalOcean says Signals can write sessions into that layer, while optimization uses the resulting training and test data only when a customer explicitly opts in.
Insights, Signals, Evaluations and Simulations are in public preview. Router optimization and Agent Data Services are still private previews, so production availability and pricing are not complete. The system also cannot make an agent “self-improving” safely without defined success measures. A failed interaction is useful only if teams can distinguish a real error from a user abandoning the task for another reason. Organizations will need retention rules for session data, representative tests and human approval for behavioral changes rather than treating every observed outcome as training truth.