By ByteBot
Every developer knows how to ship an AI agent. The harder question — what everyone quietly avoids — is what happens after it lands in production and starts failing. Silently. Frustrated users churn, tool calls break in ways that never surfaced in testing, and your only recourse is grepping logs and guessing. DigitalOcean just fixed that.
On October 12, DigitalOcean announced Agent Learning Services: a three-product feedback loop that watches your agents fail in production, surfaces the sessions worth debugging, and lets you validate a fix against those real failures before you ship it. All of this runs on top of the same Managed Agents runtime it launched in September, with no extra infrastructure to wire up.
The Improve Loop: Insights, Signals, and Evaluations
The three products are designed to chain together as one continuous loop.
Insights is the passive layer. Once enabled, it records every agent turn — token throughput, latency, error rates, cost per session, and a full trace from the model call through each tool the agent invoked. There is nothing to instrument. Insights covers Managed Agents, Serverless Inference, Batch Inference, and the Action Gateway out of the box.
Signals is where the loop gets interesting. Rather than dumping every session into a dashboard and leaving you to find the bad ones, Signals watches production traffic and automatically flags the sessions worth your attention: frustrated users, failing tool calls, timeouts, and rate limit errors. It uses rule-based detection — not another AI model stacked on top of yours — and achieves an 82% informativeness rate with a 1.52x efficiency gain over random sampling on the τ-bench benchmark. PII is redacted before processing, and traces stay within your team boundary. When Signals flags a session, you get one click to turn it into an evaluation scenario.
Evaluations and Simulations closes the loop. You take a candidate fix — a new prompt, an updated tool config, a model swap — and run it against multi-turn scenarios built directly from those flagged production sessions. The simulation runs in an isolated sandbox with the same tools and runtime harness as production, scored by an LLM judge on goal completion and other configurable metrics. You know whether the fix actually works before it touches users.
All three are in public preview today. The community tutorial repo at github.com/anishsingh20/signals-loop-lab already has a working Signals → Evaluations → deploy walkthrough if you want to see the loop end-to-end before enabling it on your own agents.
Agent Data Services: The Private Preview Complement
Running alongside the learning loop is Agent Data Services, currently in private preview. Where the learning loop focuses on what went wrong and how to fix it, Data Services handles the memory side: giving agents one managed place to read context and write back what they learn.
The core offering is Managed Data Flows — pipelines that automatically ingest, chunk, embed, and vector-index data as it arrives, whether that is multimodal content, real-time logs, or agent outputs. The underlying vector store is Weaviate, managed by DigitalOcean. Agents can query it via SQL or a direct API call, and there is a unified catalog with data lineage back to the source. You can request private preview access at try.digitalocean.com/agent-data-services-private-preview.
The Competitive Gap This Fills
AWS, Microsoft, and Google all have managed agent runtimes now. None of them have shipped an improve loop — the packaged observe-flag-test-ship cycle that most teams are doing manually today. DigitalOcean is the first cloud where you can run an agent, watch it fail, build an eval from the failure, test a fix in a prod-identical sandbox, and ship — all on one platform, all on one bill.
That matters most if you are already on Managed Agents. If you are not, it is a reasonable reason to consider it, particularly at the Agent Droplets flat pricing ($50/month covers compute, inference, and 16,000+ tools via Action Gateway).
What to Do Now
Agent Learning Services is available in the DigitalOcean control panel under the Agent Learning section. Enable Signals on your existing Managed Agents config — there is nothing additional to instrument. Flagged sessions will surface in your Insights dashboard within the first few hours of production traffic.
Agent Data Services is still private preview. If persistent agent memory across sessions matters to your use case, apply for access now. The private preview window is typically 4–6 weeks before broader availability.
One caveat worth naming: both services are previews, not GA. DigitalOcean itself frames them as “not yet hardened for production workloads.” Use them in staging or low-stakes production first while the rough edges get filed down.
