DigitalOcean Gives AI Agents a Feedback Loop—Data Services Remains Private

On October 12, 2026, DigitalOcean announced the launch of Agent Learning Services in public preview and Agent Data Services in private preview ahead of its Open Intelligence Summit in San Francisco; chief product and technology officer Vinay Kumar framed the ambition as agents that “get better with every run.” Preview offerings are supplied “as is” and are unsupported for production workloads.

The learning services connect agent traces to failure signals and then to evaluations of proposed changes. Agent Data Services supplies the managed layer for processing behavioral records and content, but admission to its private preview is separate from access to the public learning tools. The Zeniteq account distinguishes the new public-preview Simulation capability from the existing Evaluations service and notes that the data preview has no service-level agreement or support commitment.

From an agent trace to a testable change

Agent Learning Services assigns a different task to each stage of the feedback loop. Insights records an agent’s steps. Signals identifies sessions in which the agent stalled, repeated a tool call, lost a user, or failed, giving a team a specific interaction to investigate rather than a broad performance score. A flagged session can become a scenario for Evaluations & Simulations, which tests a candidate change before that change reaches users.

Signals assembles individual agent turns into sessions and highlights interactions that may need attention. Simulations can then run a proposed agent configuration through a generated, multi-turn conversation grounded in a scenario. The candidate is graded against chosen criteria, and a scenario can be retained for later regression checks. An evaluation result is evidence about that candidate under the selected scenario; it is not a measured improvement across every live conversation.

The automation boundary is narrower than the self-improvement ambition might suggest. The private-preview Inference Router optimizer can select a model based on cost, quality, or a team’s evaluations, and rerun evaluations when new models arrive. Automating the same loop for agent workflows is a future plan. Changes to those workflows currently involve human approval, so the trace-to-test sequence supports a decision rather than independently rewriting a deployed agent.

What enters Agent Data Services

The private data service handles behavioral records and multimodal assets in the same managed workflow. The Agent Data Services architecture documentation describes processing on arrival and storing derived results for retrieval, rather than running the processing models again for every query. Its two private-preview workloads are behavioral data, such as events, logs, clicks, and agent traces, and multimodal content, such as documents, images, audio, and video.

There are three documented ways to supply source material. A connected DigitalOcean Spaces or other S3-compatible bucket can trigger a data flow when a new object appears. A path in an existing bucket, including AWS S3, can be processed with read credentials without migrating the source objects. An application can also upload data directly through the API. Those are ingestion paths for an admitted workspace, not separate routes around private-preview access.

A data flow chains built-in functions in a chosen order. For behavioral input, it can assemble events into sessions, redact sensitive fields, and store a completed interaction for lookup by its identifier. For content, steps can extract text, transcribe audio, divide material into chunks, generate descriptions and embeddings, and attach categories or moderation flags. The service keeps those outputs linked to the original asset and its attributes; a video segment, for example, can remain associated with its transcript and timestamp.

Where data is stored and controlled

Each workspace contains its data flows, catalog, and query endpoint, belongs to one DigitalOcean team, and resides in one region. Structured records and assembled sessions are stored in Apache Iceberg tables, while embeddings go into a managed vector database. Callers reach both through the service’s query layer rather than accessing either store directly. Catalog lineage tracks a result back to the arrival and transformations that produced it, making a retrieved record more than an isolated answer.

Redaction can occur before a sensitive field is embedded. At query time, access rules, licensing conditions, and moderation flags determine which results a caller may receive. DigitalOcean IAM controls who can manage or query a workspace, whether the caller is a person, an application, or an agent. Data is encrypted in transit and at rest; customer data is excluded from training DigitalOcean models. These controls describe the service design during private preview, not an independently measured security outcome.

The query endpoint supports semantic and keyword search, session or record lookup, filtered retrieval with aggregations, and read-only analytical queries. It cannot create, update, or delete workspace records; data flows provide the write path. That separation matters to the agent loop: a session can be retrieved as context for investigation or evaluation, while adding fresh behavioral data follows the configured ingestion process.

What developers can access now

Insights, Signals, and Evaluations & Simulations are offered as public previews, while the Inference Router optimizer and Agent Data Services require private-preview access. Existing Evaluations was generally available before the newly added public-preview Simulation capability. Access to a public learning feature therefore does not grant use of the managed data layer described above.

An agent team can inspect and test candidate behavior with the public tools, but the private data foundation is available only to accepted participants. Broader access and explicit support terms for Agent Data Services would determine when teams can rely on that layer in production; evidence from deployed agents would show whether the combined loop improves outcomes beyond its test scenarios.

Also read:

• Cohere North 2 Gives Agents Memory—and Admins Token Caps
• DigitalOcean vs Vultr: Equal $24 Plans Deliver Unequal CPU and Storage
• THE·TEAM Buys Outreach Talent Group—The Price Stays Private