DigitalOcean Previews Agent Testing and Governed Data
Public-preview simulations and private-preview Agent Data Services offer a workflow for testing agent fixes, with important limits on access and production support.
Listen
AI narration
11:12
0:00 / 11:12
AI SummaryGenerated from this article
DigitalOcean launched Simulation as a public preview in October 2026, enabling teams to test agent changes through generated multi-turn conversations using scenario sets, simulated users, and judge models. The company also announced Agent Data Services in private preview, a managed layer for execution traces and multimodal content that applies governance controls including IAM access, field redaction, and moderation flags. However, private-preview services lack service-level agreements, support commitments, or compliance certifications, and pricing remains subject to change before general availability.
DigitalOcean’s new Simulation public preview tests agents through generated, multi-turn conversations. Builders can use scenario sets, simulated users, judge models, and conversational metrics to examine how an agent behaves as a conversation develops, beyond an isolated response.
Simulation fits into a broader post-deployment workflow: observe agent behavior, identify problematic sessions, test proposed changes, and make the underlying data available for subsequent investigation. DigitalOcean’s accompanying Agent Data Services announcement adds a managed layer for execution traces and multimodal content.
Access differs between the services. The Simulation release notes explicitly mark the feature as public preview on October 12, 2026. Agent Data Services is a private preview, available only to accepted participants. Builders can investigate the public testing capability now, but should not assume the entire workflow is equally accessible or ready for production reliance.
The Feedback Loop Starts With What an Agent Actually Did
DigitalOcean’s product framing connects three stages: Insights for observing behavior, Signals for surfacing problematic sessions, and Evaluations and Simulations for testing changes. Agent Data Services supplies a common data foundation beneath that workflow.
Collecting evidence and proving that a proposed fix works are separate tasks. An execution trace records what happened during a particular interaction, helping a team investigate the sequence of agent turns, tool calls, and intermediate results beyond the final answer.
Identifying a problematic session narrows the investigation, but it doesn’t establish the cause. A poor result might require a change to instructions, retrieval, tool behavior, or the information available to the agent. Each intervention needs a test suited to the suspected problem.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
Simulation lets teams exercise a proposed change across generated conversations before shipping it. For example, a team investigating a support agent that mishandles follow-up questions could build scenarios requiring the agent to maintain relevant context across several turns. That is an illustrative use case, not a reported DigitalOcean test result.
The Agent Data Services announcement describes agent turns, tool calls, and session history becoming structured, queryable context for the broader observability and evaluation suite. Those records give builders a concrete basis for investigating failures, proposing an intervention, and gathering evidence that it helps.
None of these functions, by itself, demonstrates an autonomously self-improving agent. Teams still need to design, assess, and control the improvement process.
Public Simulation Is New, but Evaluations Already Exists
DigitalOcean Evaluations became generally available on June 30, 2026, according to the Evaluations documentation. Simulation is the new public-preview capability added in October.
“Evaluations and Simulations” therefore should not be read as a single product with one uniform maturity level. The existing evaluation service supports test cases, evaluation datasets, selected metrics, and reusable configurations. Simulation introduces generated, multi-turn conversations.
The two approaches answer related but different questions. A dataset evaluation measures performance against a defined collection of cases. A conversational simulation can examine how behavior develops through successive exchanges, including follow-up requests and changes in conversational context.
DigitalOcean says Simulation uses scenario sets, user simulator models, judge models, and conversational metrics. The scenario set defines the situations to exercise; the simulator supplies the user side of the conversation; the judge and metrics provide the assessment.
For builders deciding where to start, the documented access split is:
Simulation: Public preview, with its capabilities described in the Evaluations documentation.
Existing DigitalOcean Evaluations: Generally available, according to the June release notes.
Agent Data Services: Private preview, requiring acceptance into the program.
The cited documentation establishes these statuses directly. For other components of the learning workflow, builders should check individual enrollment and availability details instead of extending one service’s preview label to the whole stack.
Agent Data Services Processes Context Before Retrieval
Agent Data Services makes both agent behavior and source content usable through a managed data layer, addressing a different problem from evaluation.
According to the service architecture documentation, it supports two workloads during private preview. Behavioral data includes events, logs, click events, and agent traces. Multimodal content includes documents, images, audio, and video.
Data flows process incoming material through configured steps such as transcription, chunking, embedding, redaction, field extraction, session assembly, and aggregation. The resulting records remain associated with their source and structured attributes in a catalog.
DigitalOcean says these transformations happen at ingestion. Retrieval reads stored results without rerunning the processing models for every query. This can avoid repeating the same inference work, although it does not establish a measured latency improvement or total-cost saving for a particular application.
For behavioral data, the service can assemble individual events into sessions and return an interaction by its identifier. Outputs for multimodal assets can include transcripts, scene descriptions, on-screen text, categories, moderation flags, and embeddings.
With both workloads in one catalog, a team can examine an agent’s behavior alongside the content involved in the interaction and ask whether a failure arose from the agent’s handling of information or from the information it was given.
The service supports semantic and keyword search, record lookup, filtered retrieval with aggregations, and read-only analytical queries. Queries cannot create, update, or delete workspace data; data flows are the write path. Its capabilities extend beyond document search, but it is not a general-purpose writable database interface.
Governance Travels With the Data
DigitalOcean’s governance claims cover processing and retrieval. The company says workspace access is controlled through DigitalOcean IAM, sensitive fields can be redacted or masked before embedding, and moderation flags and asset access rules are applied during queries.
Redacting a sensitive field before creating an embedding addresses a different stage of the data lifecycle from filtering a result after retrieval. Enforcing access rules when a person, application, or agent queries the workspace also places the control in the data service, without relying solely on an agent’s instructions.
The catalog records lineage: where an output originated and which transformations produced it. In an evaluation workflow, that can help distinguish a faulty source record from an extraction error or an agent-level problem.
DigitalOcean says data is encrypted in transit and at rest and is not used to train DigitalOcean models. These are vendor-described properties, not independently verified findings from this article.
Permission to retrieve a record does not make it accurate, current, or appropriate for every task. Teams still need to assess their source material and evaluate how agents use it. A controlled data layer can support that work without replacing it.
Simulated Improvement Needs a Reality Check
Generated conversations expand the range of interactions a team can test, but a simulator is not an actual user. Results depend on the scenarios chosen, the behavior generated by the user model, and the criteria applied by the judge.
A narrow scenario set may miss the behavior that caused the original failure. A judge may reward an answer that sounds satisfactory without establishing whether the agent completed the intended task. These are evaluation risks to account for, not reported defects in DigitalOcean’s implementation.
A useful testing approach would connect each simulation to an observed failure, define the desired outcome before running the test, and compare the proposed change against the current configuration. Teams should inspect representative conversations alongside aggregate metrics, particularly when a change affects tool use, permissions, or sensitive data.
DigitalOcean’s documentation does not provide an independent benchmark demonstrating that this combined workflow improves deployed agents. The announcement establishes the tools’ availability and design. Better outcomes depend on the tests and operational controls builders apply.
Preview Access Is Not Production Support
The Agent Data Services announcement carries the strongest practical caution. DigitalOcean says the private preview is provided as-is, without a service-level agreement or support commitment, and is not covered by its standard compliance certifications. Features, availability, pricing, and timing may change, and the service may be discontinued.
The Agent Data Services pricing page says there is no charge during private preview. Pricing is subject to change before general availability, so the temporary preview price is not a basis for forecasting long-term operating costs.
Builders should treat the new previews as evaluation environments, not supported production dependencies. Even the architecture documentation’s instruction to “promote” a data flow describes a configuration step; it should not be mistaken for a production-support guarantee.
The immediate opportunity is to test the public Simulation capability against a real, understood failure pattern. Teams that need the broader behavioral and multimodal data layer can request private-preview access separately and assess its governance and query behavior with suitable test data. A production commitment should wait for explicit support terms, stable availability, and evidence that the workflow improves the failures the application actually encounters.
