Skip to content

Question: is there a supported way to constrain what generated code can access at execution time? #13636

Description

@incyashraj

Duplicates

  • I have searched the existing issues

Summary 💡

Working on the execution side of agent-generated code and trying to understand AutoGPT's current model before I go further.

When an agent generates and runs code, what's the supported way today to constrain what that code can touch on the host; filesystem, network, environment? Looking through the execution path I can see the sandboxing options, but I want to understand what's considered the recommended boundary rather than inferring it.

Asking because I've been building a runtime aimed at that problem: one file that runs natively on Mac, Windows and Linux with zero permissions until explicitly granted, returning structured denials an agent can parse and adapt to rather than a generic failure. There's an MCP server in it, so an AutoGPT agent could be pointed at it as an execution target today. Repo is https://github.com/incyashraj/krate.

Two questions, either one is useful:

  1. What's the intended execution boundary for generated code in AutoGPT right now?
  2. If someone wanted to plug in an alternative execution target, what interface would it need to satisfy?

Happy to do the integration work myself if the second one has an answer.

Examples 🌈

No response

Motivation 🔦

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    Priority

    None yet

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions