DigitalOcean

When an agent gives someone a bad experience, they rarely complain. They just stop asking, and nothing in your dashboard changes. In this demo, Developer Advocate Amit Jotwani shows how to catch it: → Signals surfaces the conversation nobody reported, by watching for corrections, escalations, and tool calls that fail or loop → One click turns that session into a test scenario → Simulations replays it as a full conversation against your fix, before it reaches a user No more guessing whether the fix worked. You go from 0% to 100% and can prove it. Signals and Simulations are now in public preview: https://do.co/4yK39bk

Andi is a two-person team. Their search index covers 28 billion pages for humans and AI agents, and it runs on DigitalOcean Droplets and Spaces Object Storage. Agent signups are up 10% a day and they still haven't hired anyone to manage the infrastructure. Their infrastructure spend dropped ~90% when they migrated. https://do.co/47shj4Q

You fix one step of your agent with a prompt change, and a later step quietly breaks. A test that scores one response at a time won’t catch it. Simulations, now in public preview in DigitalOcean Evaluations, tests agents across the whole conversation: • A simulated user plays the persona you define, turn by turn • Your agent runs on the same harness and tools it uses in production, with mocks for anything that writes data or costs money • A separate LLM judge scores the conversation against your criteria, so your agent isn’t grading its own work In the demo, an invoice-checking agent passes its first test. A closer read shows its email to the vendor raised an issue that wasn’t on the invoice. One new check and three journeys later, the judge flags it. Scenarios can come from a library, a plain-language description, or, with Signals, your own production traffic. https://do.co/4yiqOP3

Open Intelligence Summit split into three tracks, and each one ended up arguing about a different layer of the open AI stack. Build: does "open" actually protect you from lock-in? Brad Gyger of Supabase offered a simple test: "Can you go pull down whatever is up on GitHub and run it yourself?" Serve: Peter Walker shared OpenRouter data showing organizations went from six models on average at the start of the year to about 13 a month. And NVIDIA's Shashank Verma pushed back on the usual framing: "A model can be open source and frontier at the same time." Run: "You have to be agnostic relative to any particular model," said Spenser Skates of Amplitude. Sam Herring of Nous Research put the principle in four phrases: "Own your weights, own your data, own your model, own your agent." Ten speakers across three tracks, and one question underneath all of it: who owns the layers your agents run on?

AI agents need to do more than process text. They need to understand documents, analyze video, and decide what to do next. NVIDIA Nemotron™ 3.5 Super VL is now available on DigitalOcean Serverless Inference, Dedicated Inference, and Managed Agents. https://do.co/47VmsTj This open multimodal model brings visual understanding, reasoning, and tool use together, making it easier to build agents that work with images, documents, and video—without stitching together separate vision and reasoning models. Hot off the heels of the Open Intelligence Summit, this is another step in NVIDIA and DigitalOcean’s shared commitment to making open AI models more accessible to developers.

So excited to have OpenHands running on DigitalOcean Managed Agents. Teams using Agent Canvas can now run coding agents in isolated cloud sessions that pause and resume with their full context, and choose DigitalOcean for agent execution, model inference, or both.

Who gets to shape the future of AI? We asked builders from OpenHands, Inferact, bolt.new, Amplitude, and Arcee AI what open intelligence means to them. Today they’re among more than 25 speakers at the first Open Intelligence Summit in San Francisco, co-hosted by DigitalOcean and NVIDIA, with the keynote streaming live at 1 PM PT. Paddy Srinivasan opens with the case for keeping the AI stack open. He’s joined on stage by DigitalOcean CPTO Vinay Kumar, with a live demo of turning an agent’s production runs into tested fixes; NVIDIA’s Adel El Hallak on open models and tools for agents; Amplitude CEO Spenser Skates on what changes when a data company’s users become agents; and Bright Data CEO Or Lenchner on giving agents the live web. Also new today: Bright Data’s search, scraping and datasets are available to agents on DigitalOcean through Action Gateway. The afternoon is built around the questions open AI teams are arguing about: where open formats actually protect you from lock-in, whether teams will run more models or fewer a year from now, when it’s worth post-training an open model, and whether you can run serious agents without depending on one model provider. Speakers include NVIDIA, Hugging Face, Nous Research, Inferact, VAST Data, Liquid AI, MiniMax, LanceDB, bolt.new, Daytona and more. The livestream links are in the comments.

One new user tried to link their bank, and the onboarding agent told them to try again later. Nothing showed up in the logs, and the agent was wrong. But that user was the first to hit a real bug. Bugs like that usually surface after they’ve hit a few hundred people. In our launch film, a team catches this one at the first user, with Agent Learning Services and Agent Data Services, launching today. Insights records every step an agent takes, with no instrumentation to set up. Signals flags sessions where users got stuck or tools failed, within minutes. Agent Data Services turns the scattered clues into one catalog agents can search: the support conversation, a session replay, a PDF notice from the bank and an old ticket nobody picked up. The team’s engineering agent finds the cause with one query. Simulations tests the fix against simulated versions of that session before it ships. Both agents in the film run on Managed Agents, where every session gets its own microVM that pauses, forks and resumes, with governed access to 20,000+ tools through Action Gateway. Agent Learning Services (Insights, Signals, and Simulations) is in public preview today. Agent Data Services is in private preview, and you can request access. Links are in the comments.

“A model shouldn’t earn its keep because it looked good in a demo or passed some benchmark. You want it to validate itself on real data.” Vinay Goel, engineer at Amplitude, on how they built Wave: a new product that reads your product data, decides what to build, ships it, and measures whether it worked. The data science agents behind it run on DigitalOcean Managed Agents. Each run gets its own session, governed access to the tools it needs, and open models when they fit the job. @Amplitude’s engineers stay focused on the product. Spenser Skates, Amplitude’s CEO, on stage at Open Intelligence Summit this Tuesday. https://do.co/4hVe8aN