DigitalOcean

A paused agent session resumes in 305 ms (p50), and the agent responds in 2.43s, versus 2.47s for one that never paused. @salman_paracha on how Managed Agents is built for a workload that’s mostly idle.

"Building an AI product goes beyond just selecting a model. It’s a systems problem.”

@vinaygo of @Amplitude_HQ on how they built Wave, a proactive agent that finds what’s worth building and hands you ready-to-merge PRs. All the time. The data science agents behind it run on DigitalOcean Managed Agents.

Hear more from @Spenserskates at Open Intelligence Summit on Tuesday. Keynote streams live: do.co/4hVe8aN

I’m a product person at @digitalocean and I want to hear from developers: what are the most important features in a cloud platform for you?
Every reply gets read. Help me make it better 🫶

The Open Intelligence Summit agenda is live, with new speakers from @NousResearch, @liquidai, Bright Data, @fivetran, @dbt_labs, @VAST_Data, @supabase, and Premise. Open models, data stacks, and agents. do.co/4zlAb1k

yeah AI agents are no longer just chats anymore.

Imagine a sandbox agent opening cloud files with an ordinary file path.

That's Box Mount.

Box Mount gives sandbox agents ordinary file paths to cloud files in @Box without building a download and upload workflow.

In this demo, an agent on @digitalocean reads deal room documents locally and writes a diligence memo back to the cloud.

All through Box Mount.

Elements is a full build stack for web apps. In this thread we'll deploy a real help desk app: two load balanced @digitalocean droplets + managed Postgres. 40s initial deploy, 1s hot deploy. Spicy take at the end...

“Owning your intelligence is owning your future.”
@latkins, CTO of @arcee_ai, on open weights and everything built on top of them.

Open Intelligence Summit keynote streams live Oct 13: do.co/4zo4nZS

Your support AI agent said it refunded $40, but in reality, it refunded only $0.40. Dashboards called it healthy.

Signals (private preview) on DigitalOcean Managed Agents reads your agent's real conversations and flags the ones that went wrong.

Personal details are hidden, the data stays in your region, and it is never used to train AI for others. Full test and logs: do.co/474YmoX

LLM Routing, clearly explained!

(routing can cost more than not routing at all)

The basic idea sounds simple:

Send easy requests to cheaper models and reserve frontier models for difficult tasks.

But production routing is rarely that straightforward.

Every classification adds latency. General-purpose models can misread intent. Routing rules become outdated. And switching models during an agent session can destroy prefix-cache savings, increase costs, and create inconsistent behavior.

In this video, I cover:

→ Why DIY routing often breaks in production
→ How prefix caching changes the economics
→ Why agents need session pinning and model affinity
→ How Plano and Arch-Router handle intent resolution
→ How to build, test, and evaluate a router with DigitalOcean

Routing works, but only when the infrastructure accounts for the full session, not just the latest request.

The real question is not whether routing can reduce costs.

It is how much of your AI traffic has been overpaying this whole time.

Get started with Digital Ocean inference router: do.co/4i8JBrO

I recently wrote a detailed article on the same topic. The article is quoted below.

@nvidia is co-hosting the Open Intelligence Summit.

@JensenHuang's letter in June helped start the open intelligence movement. On October 13, we're bringing the movement to The Midway in San Francisco.

Two NVIDIA speakers are on the agenda:

A sweet treat to get you through SF Tech Week.

Our SomiSomi pop-up is available all week. Be sure to check out flavors inspired by our friends at @MiniMax_AI, @arcee_ai, and @weaviate_io.

.@kevinvangundy at SF @techweek: over the long run, open source will be the dominant modality for AI. His advice for builders today is to go multi-harness and multi-agent, and to lean hard on evals and iteration speed so you can swap models as they leapfrog each other.

Hop on, we're going to SF Tech Week.

All week: Register for our free to ride, branded shuttle bus. do.co/3Ub8yJu

Thursday: Intelligence in the Open with @MiniMax_AI. Expect short research talks, open discussions, and snacks. do.co/4yCgTFa

We usually move workloads. This week, we’re moving builders.

@digitalocean’s FREE #SFTechWeek shuttles run Oct 5–9, 8am–8pm. Hop on, hop off, get to your next big idea.

Find your ride on the live map:
digitaloceanevents.com

Drop a file in the deal room, and a sandbox agent wakes up and rebuilds the review memo with the latest info. @chriskim_dev built it with @Box and DigitalOcean Managed Agents.

"If you commit to one particular lab, there's a good chance you're going to be left behind."

@OpenHandsDev CEO @rbren_dev on choice and owning your intelligence. Hear more from him at the Open Intelligence Summit, Oct 13 in SF: do.co/4ArVTlL

Your TTFT benchmark is only as good as your stream instrumentation.

Reasoning models can emit reasoning_content before content. Watch only content, and you can miss tokens the model is already generating.

My takeaway: measure first streamed token and first answer token separately. Then inspect tail latency at the context lengths your agents actually use.

Those metrics answer different questions: when inference starts responding, and when your application can act.

@digitalocean’s audit 👇
digitalocean.com/community/tuto…

You're probably overpaying for AI in your app.

Most apps hardcode one model. Fixing a typo and refactoring an entire codebase hit the same frontier model at the same price.

DigitalOcean's Inference Router fixes this with a one-line change:

Before → model = "gpt-oss-120b"
After → model = "router:writing"

The router reads every request, figures out what kind of task it is, and sends it to the best-fit model.

You pick the strategy:

- Cost: the cheapest model that can still do the job
- Speed: the fastest time to first token
- Optimal: DigitalOcean's own recommendation

You also get:

- Automatic fallback when a model is down or rate limited
- Cache-aware routing so related requests stay on the same model
- Presets for general tasks, software engineering, writing, and documents

I tested it with the writing preset. A one sentence grammar fix and a full post went through the same router, and it classified them as two different jobs. I didn't write a single routing rule.

The trade-off: around 200 ms of extra latency per request.

Lawo, a legal tech company running 130+ agents and 500M+ tokens a week, reports cutting its inference cost by 40% after switching.

Nobody wants to maintain routing logic that breaks every time a new model drops.

Link: do.co/4hc06lS

Thanks to @DigitalOcean for partnering with me on this post.

Excited to take the stage at the Open Intelligence Summit with @digitalocean and @nvidia!

Hope to see you there.

Getting open models production-ready is mostly software tuning. @DebarshiRaha talks it through with @VictorSuOrtiz and the vLLM team at Open Intelligence Summit on Oct 13.