Pull to refresh
Logo
DigitalOcean launches Managed Agents public preview and Batch Inference

DigitalOcean launches Managed Agents public preview and Batch Inference

New Capabilities

Agents get isolated runtimes, brokered credentials and a monitoring layer, while inference jobs get cheaper to re-run

Overview

Updated Oct 29

DigitalOcean's Managed Agents, in public preview since September 22, moved a real production backend from a phone in plain chat: 31 tables and 145,432 rows, with no cloud credential ever in the agent's hands. Each session runs in its own isolated virtual machine, and a gateway attaches tool credentials only at the moment of a call.

DigitalOcean also added Batch Inference in October, running large jobs asynchronously at up to half the real-time rate. In late October it opened Signals, which flags where a user got stuck or a tool kept failing in a production agent session. It also released Cuttlefish, an iPhone app for approving or denying an agent's tool calls mid-session.

Why it matters

Teams can now run production AI agents without building the infrastructure themselves — and cut large inference job costs by up to half.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

16,000+
Tools on Action Gateway
One managed endpoint connects agents to tools from more than 500 providers, as of October 2026.
50%
Batch Inference cost cut
Asynchronous workloads cost up to half the rate of real-time inference on the same engine.
90%
Cached input discount
Prompt caching is generally available on Serverless Inference, billed at up to 90% off the context an agent resends.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Organizations Involved

Timeline

September 2026 October 2026

8 events Latest: October 28th, 2026 · 2 weeks ago
Tap a bar to jump to that date
  1. DigitalOcean Signals opens in public preview

    Latest Launch

    Signals reads production conversations on Managed Agents and Inference Engine, with consent, and turns them into 25 signal types pointing at the exact step behind a failure. It is off by default and free during preview, and flagged sessions can be exported as datasets.

  2. Cuttlefish iOS app brings Harness Runtime sessions to the iPhone

    Launch

    The app, in public preview on the App Store, lets users follow and steer sessions, approve or deny tool calls, and work with workspace files and checkpoints.

  3. Managed Agents enforces OpenShell boundaries; Batch Inference debuts

    Launch

    Production agents now run under NVIDIA OpenShell policy profiles that lock the filesystem and allowlist the network. Inference Engine adds asynchronous batch processing at up to 50% lower cost.

  4. Harness Runtime adds bare sandboxes and renames the adapter field

    Product Update

    Setting `harness: none` boots a base image with a general-purpose toolchain and no agent running, managed through `doctl` and the API. The environment spec field `agent` is now `harness`; the old name is still accepted.

  5. Three OpenAI models retired from DigitalOcean Inference

    Product Update

    o1, o3-mini and GPT Image 1 are deprecated, with DigitalOcean directing users to GPT-5.6 Sol and GPT Image 2.

  6. Harness Runtime templates patched against an RCE flaw

    Security

    DigitalOcean rebuilt its Harness Runtime platform templates to address CVE-2026-111497, a remote code execution vulnerability. Sessions created after the update use the patched templates automatically.

  7. Prompt caching reaches general availability on Serverless Inference

    Product Update

    Cached input is billed at up to 90% off and time to first token drops by up to 80%, with no change to request code across supported models in the open catalog.

  8. DigitalOcean opens Managed Agents public preview

    Launch

    Managed Agents opens to everyone, pairing per-session Firecracker microVMs with an Action Gateway endpoint to 16,000+ tools.

Scenarios

1

Managed Agents reaches general availability

Likely Resolves by Q2 2027

Discussed by: DigitalOcean's product blog and preview terms

The preview terms warn features may change or be withdrawn before general availability. If the OpenShell enforcement holds up in production — the huffl migration suggests it does — DigitalOcean moves Managed Agents out of preview, likely with pricing tiers tied to the new Agent Droplets bundling.

2

Cloud giants ship rival managed agent runtimes

Possible Resolves by End of 2027

Discussed by: Platform observers tracking AWS, Google Cloud, and Microsoft

Managed agents with brokered credentials are the obvious next cloud battleground. A hyperscaler publicly launching a comparable runtime would put DigitalOcean's preview under direct competitive pressure, testing whether its OpenShell posture and tool gateway can hold a niche against larger platforms.

3

Credential brokering survives a public security test

Uncertain Resolves by End of 2027

Discussed by: Security researchers reviewing Action Gateway

The whole design leans on credentials never reaching the model or sandbox. A public penetration test, bug bounty finding, or breach disclosure on Action Gateway would validate or break that claim. A clean external audit strengthens the path to general availability; a leak of production credentials would force a redesign and erode trust in the platform.

4

Signals exports become the main way teams evaluate their agents

Possible Resolves by Apr 30, 2027

Discussed by: DigitalOcean's Signals announcement

Signals already lets flagged sessions leave the dashboard as datasets. If teams feed those exports into Evaluations and Simulations instead of building their own test sets, DigitalOcean's monitoring layer becomes the front door to agent testing as well as production watching.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

November 2014

AWS Lambda launch (2014)

Amazon launched a compute service that runs code in response to events and bills per invocation, with no servers to provision or patch. Developers uploaded functions and AWS handled the rest.

Then

Developers shifted event-driven workloads off always-on servers and paid only when code ran.

Now

Serverless became a standard deployment model, abstracting away VM management for a generation of builders.

Why this matters now

Just as Lambda abstracted server management, Managed Agents abstracts agent infrastructure — microVMs, session persistence, and governed tool access — so teams ship agents instead of plumbing.

June 2021

GitHub Copilot launch (2021)

GitHub launched an AI pair programmer that suggests code inline in the editor, built on OpenAI's Codex model. It reached millions of developers within its first year.

Then

Developers adopted AI-assisted coding at scale, normalizing models inside the daily workflow.

Now

Copilot made AI a suggestion engine; later tools pushed it toward execution, where agents act on systems rather than propose edits.

Why this matters now

Copilot delivered AI assistance at the keyboard. Managed Agents takes the next step, letting a model execute production migrations and infrastructure changes with guarded credentials and enforced sandbox boundaries.

Sources

(10)