---
schema_version: "newruntime-agent-readable-v0.1"
type: "post"
slug: "devin-outposts-agent-execution-plane"
title: "Devin Outposts Splits the Agent Brain from the Execution Plane"
description: "Devin Outposts moves command execution, repository access, and sandbox lifecycle into customer-controlled infrastructure while the agent loop remains in Cognition's cloud."
status: "published"
published_at: "2026-07-22"
topics: ["coding-agents","agent-runtime","infrastructure"]
source_urls: ["https://x.com/cognition/status/2079612226252726615","https://developers.cloudflare.com/sandbox/tutorials/devin-outposts/","https://modal.com/blog/devin-outposts-run-devin-in-modal-sandoxes","https://www.daytona.io/docs/en/guides/devin/devin-outposts/"]
routes: {"html":"https://newruntime.com/posts/devin-outposts-agent-execution-plane/","markdown":"https://newruntime.com/posts/devin-outposts-agent-execution-plane.md","json":"https://newruntime.com/posts/devin-outposts-agent-execution-plane.json"}
source_format: "markdown"
---

# Devin Outposts Splits the Agent Brain from the Execution Plane

Devin Outposts is the same basic product move as private runners, self-hosted build agents, and VPC connectors, but for autonomous software work. The agent brain can stay hosted while the execution machine moves into infrastructure the customer controls.

Cognition's public post frames it directly: run Devin on a Mac mini, GPU box, private VM, or Kubernetes cluster next to internal services. Cloudflare, Modal, and Daytona each describe partner implementations where sessions run in isolated sandboxes, containers, or owned compute while Devin still manages the agent loop.

## What changed

This splits an agent product into two planes:

- Reasoning plane: planning, model calls, tool selection, and conversation stay with Devin.
- Execution plane: shell commands, file edits, repository access, browser automation, checkpoints, and machine lifecycle run where the customer chooses.

That matters because many real engineering tasks are blocked by environment, not by model intelligence. A data task may need private Snowflake access. An ML task may need an H100. A regulated codebase may not be allowed to leave a controlled network.

## New Runtime Read

Coding agents are starting to look like managed control planes plus bring-your-own execution fleets.

The product boundary is no longer "does the agent have a VM?" It is who owns the VM, where secrets live, how checkpoints persist, what traffic is allowed, and whether the agent can reach the same internal systems as a human engineer without exporting the whole codebase to a vendor sandbox.