Alex' Gardenアレックスの庭

2 min read

AI file sharing — agent and sub-agent invites (2026-05-10, 10:04)

Mid-morning one-liner addition to the Kumo design: the agent-member model should support agents inviting other agents (sub-agents), with those sub-agents acting as real org users. Extends the “humans and agents share one auth system” pillar of Kumo to recursive provisioning — an agent with the right scope can spawn other agents under it.

Voice Transcript

Recorded: May 10, 2026 at 10:04 AM

Reflecting on this AI native file sharing, uh, cloud stuff, I guess it should also have the possibility to invite agent and create sub-agents basically inside, uh, and they act as real users

Where this fits Kumo

  • The current Kumo design models humans and agents as the same members table with user_type (human/agent) and auth_type (password+session / api_key) columns. Invite flow vs. provision flow are the same scope/permission UI with a different credential at the end.
  • This memo adds a third actor: agents provisioning agents. If an agent has the right scope (analogous to a human admin), it can call the same POST /members/agents endpoint to spin up a sub-agent with a scoped API key — same flow as a human admin, just performed by an agent over MCP/CLI.
  • “They act as real users” — the sub-agent gets a member_id, shows up in the org’s members list, and obeys the same path-glob verifier. No special-casing.

Open questions

  • Should sub-agent scopes be bounded by the parent agent’s scope, or can a sub-agent be granted scopes outside its parent’s? (Default: bounded — agents can only delegate what they have.)
  • Audit trail — does the action log carry the (parent_member_id → child_member_id) chain, so a misuse can be traced back to the human who provisioned the top-level agent?
  • Sub-agent revocation — does revoking a parent agent automatically revoke its descendants? (Default: yes; tree-revoke.)

Graph View

Backlinks