agent-identity-access-model

name	agent-identity-access-model
description	Configure and roll out workspace-owned “agent identities” for team-wide AI agents with least privilege, strong boundaries, and reliable auditing.

Instructions

Goal

Set up an access model for agents operating in shared team spaces (channels/projects/workspaces) where tool access is:

• Owned by the workspace (not a single person)
• Scoped to a shared space (e.g., a channel)
• Centrally revocable
• Auditable end-to-end

Core concepts to implement

1. Use workspace-owned tool identities

   • Provision service accounts / app identities for the agent, created by an admin.
   • Do not rely on individual users’ personal accounts for shared-channel automation.

2. Scope permissions to the shared space

   • Treat each channel (or equivalent shared space) as the unit of authorization.
   • Ensure removing the identity ends access everywhere the identity was used.

3. Roll out access gradually

   • Start with a baseline profile in a small number of channels.
   • Review audit trails to confirm the access is used as intended.
   • Expand access deliberately, one grant at a time.

4. Maintain hard boundaries

   • Shared spaces must not become a side door into private documents.
   • Ensure memory and access follow identity boundaries (private stays private).
   • Keep boundaries firm enough that access never travels somewhere it wasn’t granted.

5. Prefer a safe credential injection pattern

   • Store credentials independently.
   • Map credentials to the channel’s identity.
   • Inject credentials at the network boundary at request time.

6. Constrain and audit external activity

   • Block outbound requests to hosts the admin has not explicitly allowed.
   • Record an audit trail for every routine, memory write, and network call made using agent credentials.
   • Ensure connected systems also receive their own action logs when applicable.

Examples

Example: staged rollout checklist

• Choose 2–3 low-risk channels for initial deployment.
• Provision least-privileged tool identities for the baseline tasks.
• Enable access only for those channels.
• Observe the audit trail over real work.
• Tighten boundaries if any access appears broader than intended.
• Expand to additional channels only when the work clearly justifies it.

Example: boundary tests

• Verify that content learned in a private channel does not appear in other channels.
• Verify that disabling the identity removes access across all integrations that used it.
• Verify outbound network requests are blocked for unapproved hosts.

Source

• https://claude.com/blog/agent-identity-access-model