Skip to content

Agent isolation via micro VMs #17

Description

@rlnorthcutt

Summary

Run agents and sub-agents in micro VMs (FaaS-style) for stronger isolation than the current process/container-level boundary.

Motivation

As agents get more autonomy (file access, tool use, custom apps), process-level isolation is a thinner boundary than we'd want for a tool that markets itself on running safely on the user's own hardware.

Proposed design

Each agent/sub-agent run gets its own micro VM. The system caps the number of concurrently running agents based on available resources, and queues additional requests rather than overloading the host.

Alternatives considered

Current process/container isolation (status quo); heavier per-agent full VMs (rejected for resource cost — hence micro VMs specifically).

Risks, drawbacks, and open questions

Resource overhead per micro VM; cold-start latency; how the concurrency cap and queue interact with a user's expectation of running several decks/conversations at once; host resource detection needs to be reliable for the cap to mean anything.

Is this a breaking change?

Unsure

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    roadmapTier 1 item — triggers auto-add to the org project

    Type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions