Let AI agents do the work — without handing them a master key
A licensed financial institution’s digital-workforce programme
What it looked like before
The business wanted AI agents to take over customer enquiries and first-pass document review, but compliance had a hard line: an agent must never touch production data with a developer’s personal credentials, and every data access must trace back to “which user authorized which agent”. The earlier pilot had all agents sharing one service account, so when something went wrong there was no one to hold accountable.
What we delivered
Built on UIAM’s Agent IAM: agents register as MACHINE principals with their own credentials and roles; acting on a user’s behalf goes through RFC 8693 token exchange, with permissions computed as the intersection of application and user permissions; every tool call is checked against authorization and quota, one call at a time. Uniclaw serves as the agent runtime, connected to WeCom and internal systems.
- An agent is a MACHINE principal — it no longer impersonates a human user, and its token carries principal_type=machine
- The delegation chain is expressed with the act claim — “who acts for whom” — nestable across hops and fully replayable
- Tool-level authorization lives in code annotations, not in documents people have to remember
- Compliance can pull an agent’s complete call record at any time, including which permissions were used
The compliance review passed in one round. The key was not how well we presented it — it is that every item can be looked up in the audit records.
Hand the complexity of identity, agents and private domain to one governable kernel
Whether you are replacing an existing IAM, building an agent platform, or trying to make private-domain operations actually work — start with a 30-minute architecture call. We will first judge whether this is the kind of problem we are good at, and say so plainly if it is not.