AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 26, 2026

DigitalOcean announced public preview availability for Learning Services and Agent Data Services, two offerings meant to let developers build, evaluate and keep improving AI agents on one AI-native cloud platform. InfoQ's Sergio De Simone reported the launch. The pieces sit on top of the company's existing Inference Engine and Managed Agents, so the platform story is that agents can think, act and now improve, using the customer's own data.

The public preview covers Insights, Signals and Evaluations, which includes Simulations. Two other items are more restricted: Inference Router optimization and Agent Data Services are in private preview, meaning access is limited rather than open to all customers.

Learning Services bundles three capabilities. Insights records what an agent does. Signals looks across that activity to spot failures and performance gaps. Simulations then tests candidate fixes by running realistic conversations before anything reaches production. Read together, the loop is capture, diagnose, rehearse: record real sessions, find where they break, then replay improved behavior against simulated users rather than live ones.

The Inference Router, though separate from Learning Services, is central to the continuous-improvement pitch. It picks models based on what a task requires, trades cost against quality, and reruns the same evaluations when new model versions appear. That last part matters in practice: evaluation suites age quickly as model lineups churn, so tying evals to the router gives teams a way to re-check behavior after a model swap without rebuilding the harness by hand.

Agent Data Services supplies the managed data layer underneath. It processes and indexes documents, video, audio, logs and agent sessions, and exposes them through a unified catalog with built-in governance and access controls. The connection to Learning Services runs both ways: Signals writes sessions into the data layer, and Optimizer draws training and test data from it, but only where the customer expressly opts in. That opt-in boundary is the detail worth noting for teams with sensitive session data, since it makes consent the gate on using production traces for training.

Reaction from practitioners focused on that write-back path. Jonathan Saring, co-founder of LakeOps, called it the most interesting part, noting that once traces, tool calls and outcomes feed evaluation or learning they stop being just logs and become production data. Deepanshu Udhwani, a senior engineer, argued the write side is underrated next to the retrieval side most teams discuss, and that many teams build strong retrieval but log outcomes poorly, leaving no way months later to tell which tool calls actually worked. Keith Zhai, co-founder of TinyFish, framed Signals and Simulations as a way to turn one failed user session into a fix before hundreds more users hit the same problem, adding that healthy tool calls alongside abandoned users is where agent stacks tend to break.

This suggests the product is aimed at a familiar gap. Agent observability tooling has largely been about traces and dashboards, while evaluation has lived in separate frameworks and scripts. Folding capture, diagnosis, simulation and a governed data store into one vendor platform reduces integration work, but it also concentrates session data and eval logic in a single place — a trade-off teams with existing observability stacks will need to weigh. The opt-in requirement for training data is a sensible default, and a likely differentiator against tools that treat production traces as fair game.

Availability is split, so the practical starting point for most teams is Insights, Signals and Simulations in public preview, with the router optimization and data services to follow for those admitted to private preview.

Why it matters: teams running agents in production get a managed path from recorded sessions to diagnosed failures to pre-deployment simulation, rather than stitching those stages together themselves. The split preview means the evaluation loop is testable now while the data foundation and router tuning remain gated. For product teams, that makes the write-back of session outcomes — and the opt-in control over using it for training — the parts to evaluate first, since that is where production traces stop being passive logs.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

DigitalOcean has announced the public preview of Learning Services and Agent Data Services, enabling developers to build, evaluate, and continuously improve AI agents within a unified, AI-native, cloud platform. By Sergio De Simone