DigitalOcean Launches Learning and Data Services to Enable Self-Improving AI Agents

DigitalOcean has launched Learning Services and Agent Data Services to empower developers to create, evaluate, and continuously enhance AI agents on a unified cloud platform. These new services complement DigitalOcean's Inference Engine and Managed Agents, allowing agents to learn from real-world interactions, pinpoint weaknesses, test improvements, and utilize company data.

Learning Services comprises three components: Insights, which logs agent activity; Signals, which detects failures and performance issues; and Simulations, which evaluates potential enhancements via simulated conversations before actual deployment. Inference Router, although distinct from Learning Services, plays a pivotal role in DigitalOcean's approach to refining agents. This component optimizes model selection based on task requirements, balances cost and quality, and re-assesses new model versions.

Agent Data Services offers a secure, managed data infrastructure for AI agents. It processes and indexes diverse content such as documents, video, audio, logs, and agent sessions, rendering this information searchable and accessible through a unified catalog with built-in governance and access controls. Additionally, Agent Data Services collaborates with Learning Services to facilitate ongoing agent enhancement: Signals archives sessions into Agent Data Services, while Optimizer draws its training and test data from the same repository upon customer approval.

In a LinkedIn post, LakeOps co-founder Jonathan Saring highlighted the significance of the write-back loop, which transforms logs into production data by incorporating traces, tool calls, and outcomes into evaluation or learning processes. Senior engineer Deepanshu Udhwani emphasized the importance of tracking outcomes, suggesting that many teams excel at retrieval but often overlook logging proper results.

This oversight can hinder the ability to discern which tool calls were effective versus those that merely appeared functional.

Keith Zhai, TinyFish co-founder, underscored how Signals and Simulations can rectify failed user sessions before numerous other users encounter the same obstacles. He noted that the bottleneck often lies in the team's ability to read and understand the actual user experience, as opposed to merely storing session data. The public preview of DigitalOcean Insights, Signals, and Evaluations, as well as Inference Router optimization and Agent Data Services, is now available.