DigitalOcean previews a learning loop, with automatic improvement in routing first

On 12 October 2026, DigitalOcean announced Agent Learning Services and Agent Data Services. The press release headline says the products are for “recursively self-improving” agents. The same release then limits what is automatic today.

What is in preview

Agent Learning Services, which the release lists as Insights, Signals, and Evaluations and Simulations, are in public preview. The Inference Router optimizer and Agent Data Services are in private preview. The release says teams can request data-services access, and that DigitalOcean will demonstrate the services at its Open Intelligence Summit on 13 October in San Francisco.

The release’s own caveat says preview services are provided as is, are not supported for production workloads, and may change, never become generally available, or be discontinued. The data-services post adds that the private preview has no service-level agreement or support commitment and is not covered by DigitalOcean’s standard compliance certifications. None of the launch pages states a price.

What “self-improving” means in the release

The release describes a loop. Insights records what an agent does. Signals flags sessions “where the agent stalled, looped, lost the user, or failed.” Evaluations and Simulations test a change on realistic conversations built from flagged sessions, before customers see it.

It then says the recursive part is available today for models. The Inference Router picks the best-scoring model for each task, optimized for cost, quality, or the team’s own evaluations, and re-runs those evaluations as new models ship. “Next,” the release says, DigitalOcean plans to bring the same automated loop to agent workflows, “augmenting the current human approvals.” So the automated workflow loop is a plan. Humans approve workflow changes now. The routing optimizer is the piece the company says is already choosing models from scores.

Signals, in DigitalOcean’s own wording

The Learning Services page says Signals watches traffic on Managed Agents and Serverless Inference and “flags the sessions that went wrong, with nothing extra to instrument.” The Signals post, which shows an update time of 11 October 2026, says the same thing as “There’s nothing extra to instrument,” because Harness Runtime and the Action Gateway already emit events. The product page classifies sessions as Interaction (user sentiment, corrections, escalations, quits), Execution (invalid tool arguments, retry loops), and Environmental (timeouts, rate limits, context overflow). The release’s shorter list, stalled, looped, lost the user, or failed, is the press wording. The page says Signals is opt-in, that personal data is scrubbed, and that traces stay inside the customer’s DigitalOcean team and are not used to train shared models.

The product page also draws a line the press release does not. The release folds Insights into Agent Learning Services. The product page’s FAQ describes Insights as end-to-end tracing, metrics, logs, and traces for each agent turn, and describes Learning Services as Signals plus Evaluations and Simulations, the step that uses that record. Both pages are DigitalOcean’s. They do not use one name for one product.

Simulations, the company says, replays multi-turn conversations in an isolated sandbox on the same harness and tools, with mocks for tools that write data or cost money. The Signals post says third-party agents such as LangChain, Hermes, and CrewAI are not supported in this preview, and that Claude Code, Codex, and OpenCode are among the harnesses Harness Runtime supports.

The data catalog

Agent Data Services, in private preview, is the store the release says the loop runs on. The release says video is transcribed and described, documents are read and indexed, and logs and agent sessions are assembled with configurable redaction controls, “searchable by meaning in a single call.” It says the result lands in one catalog, “designed to be logically isolated to that customer.” It says the Optimizer draws training and test data from that catalog only where the customer expressly opts in. The data post says the catalog is queryable by SQL, vector search, or API. Those capabilities are the company’s description of a private preview, not a result from an outside test.