Skip to main content
An agent that runs in the cloud goes through the same loop as one on your laptop. What differs is where it runs, and so where Helix reads its traces and where an approved fix lands.

​
Mutagent managed agents

A managed agent is an agent.md folder (the agent’s prompt, tools and skills) that Mutagent runs for you in a cloud sandbox, with the LLM providers and Environment of your workspace:
A cloud run does not upload files from your machine. Put the text in the prompt, or package the files the agent needs under harness.files in the agent folder. See Managed agents and the managed agents quickstart.

​
Agents on other platforms

Mutagent does not deploy to these platforms. When your agent already runs on one of them, Helix evaluates and diagnoses it from the traces the platform sends to your observability backend, and ships approved fixes as pull requests on the agent’s source; your own deploy pipeline takes it from there.

Paperclip

An agent platform of its own, not a Mutagent product. Agents you run there are improved through their traces and a pull request on their source.

Mastra Cloud

Mastra agents are a supported code target: fixes land as a pull request on the Mastra project you deploy from.

Vercel

Agents deployed on Vercel are improved through their traces and a pull request on the repository Vercel builds from.

Claude Agent SDK agents

Claude Agent SDK agents are a supported code target: fixes land as a pull request on the agent’s code.

​
How it fits the loop

The shape does not change: Evaluate scores the agent, Diagnose finds why it fails, and Optimize applies the fix you approve. For a managed agent, the updated agent.md folder is deployed again as a new revision; for an agent on another platform, the merged pull request goes out through that platform’s deploy.