Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Baseten Sandboxes keep agents running for days

Written November 7, 2026

Baseten Sandboxes give agents microVMs that snapshot filesystem, memory, and processes — and resume in 25ms. Free until Nov 16. The limits that matter.

What it is

Baseten Sandboxes, announced November 5, 2026, give an agent its own persistent computer. Each sandbox is an isolated environment on its own microVM, with its own kernel and hardware-level isolation, so an agent can run untrusted code, change files, and run processes without lateral movement to anything else. When the agent disconnects, everything is snapshotted — filesystem, memory, and actively running processes — and compute scales to zero while the environment sleeps. When work resumes, the sandbox restarts in about 25 milliseconds, up to five times faster than other sandboxing tools by Baseten's own measure. Baseten says it can boot a sandbox from standby in 25 milliseconds and scale to thousands of sandboxes within seconds.

The point is the long task. An agent writing and running a website's code, or downloading sources, analyzing them, and collecting research results, keeps its working state between steps instead of paying the setup cost again on every restart. The sandbox can pause for human input and continue with its environment intact. The built-in dashboard shows live metrics for API requests, error rate, CPU, and memory, plus the sandbox's region, RAM, image, and labels, so the environment's activity is visible without a separate monitoring setup.

Each sandbox starts from an image — the operating system, language runtimes, tools, and starting files — either a built-in image maintained by Baseten or a custom image built from a Dockerfile. Every successful build creates a new tagged version, and each new sandbox starts from a copy of one version; files a sandbox creates or changes while it runs do not modify the image. Sandboxes are worked with through the Management API, the sandbox runtime API, the Baseten CLI (baseten sandbox exec for commands, baseten sandbox connect for a terminal), SDKs, or MCP. The docs set a default limit of 10 concurrent sandboxes per organization, with support able to raise it.

What it costs

Sandboxes are free to use until November 16, 2026. No price after that date is published in the announcement or the docs: the free window ends, and the cost of running sandboxes past it is unstated. Treat November 16 as a decision date, not a deadline to rush past — a long-running agent started on free sandboxes needs a home after the window closes, and the price of that home is not yet public.

What to do first

  1. Give one agent one sandbox. Create a sandbox per user or per project, not one shared sandbox for everything. The isolation boundary is the unit that keeps one agent's mess out of another's work.
  2. Put results outside the sandbox. Deleting a sandbox, including deletion when it expires, permanently removes its local data. Anything that must survive — finished code, research results, reports — goes to storage outside the environment or onto a mounted sandbox volume before the sandbox is deleted.
  3. Let it sleep. The suspend-and-resume design is the cost control: disconnect when the agent is done for now, and the compute bill stops while the state waits. A sandbox left running for no reason is the one cost the design cannot save you from.
  4. Plan past November 16. Before the free window ends, decide whether the sandbox stays and on what terms. A long task that spans the cutoff needs its price known before it crosses it, not after.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: OpenAI's Agents API: the harness is now hosted · November shutdowns: Runloop, Conductor, Moondream · Check Copilot cloud sandbox billing first