DigitalOcean Agent Learning Services: keep the human gate

DigitalOcean wants your AI agents to get better every time they run. On October 12 the company announced Agent Learning and Data Services, a set of tools that record what an agent does, flag where it fails, and test a fix before customers see it. The release calls them "recursively self-improving" agents. Read the fine print: today, the loop that changes agent workflows still waits for a human to approve. That is the part you should keep.

What actually happened

Per the release, syndicated through Business Wire ahead of DigitalOcean's Open Intelligence Summit on October 13, Agent Learning Services is three products:

• Insights records each step an agent takes.
• Signals flags sessions where the agent stalled, looped, lost the user or failed.
• Evaluations & Simulations builds realistic test conversations from those flagged sessions and runs a proposed change against them before it goes live.

Those three are in public preview. Two more pieces are in private preview: an Inference Router optimizer, which picks the best-scoring model per task for cost, quality or your own evals and reruns the evals when new models ship, and Agent Data Services, a managed data layer that indexes documents, video and agent logs, with configurable redaction. DigitalOcean says its optimizer only trains or tests on that data when a customer opts in.

The automatic loop works on model choice today. For agent workflows, DigitalOcean says it plans to extend the loop to "augment" the current human approvals. No pricing was given. The release also says preview services are "as is" and not supported for production workloads.

This builds on Agent Droplets, the $50-a-month agent hosting plan DigitalOcean launched on October 1.

Why agent evaluation matters for your business

Most small-business agents fail quietly. A support bot that loops on a refund question does not throw an error. It just loses the customer. A tool that flags stalls, loops and drop-offs is the piece most teams skip, and it is the piece that tells you whether the agent is earning its cost.

"Self-improving" is an eval problem, not a magic trick. An agent only gets better if it is scored against tests that match your real customers. Build those tests from your own failed sessions, which is what Signals plus Simulations does. Generic benchmarks will not tell you if the bot handles your return policy.

Let it pick models. Do not let it rewrite the prompt unsupervised. Automatic model routing against your own evals is low risk: the eval set is the guardrail. Automatic changes to agent instructions and workflows are different. One bad change can promise a discount you do not offer. Keep a person approving those changes, and keep a changelog you can roll back.

Watch where the logs go. Session logs hold customer names, orders and addresses. Turn on redaction before you send them to any vendor's data layer, and leave training opt-in off until you have read the terms.

Sources: DigitalOcean announcement via AIwire / Business Wire, StockTitan release copy, DigitalOcean evaluation docs.