Capability

Agents with their own identity and their own limits

On Substrait an agent is a tool its builder owns, governed like every other tool and then some. Each agent has its own identity, authority narrower than its owner's, a check on every action as it happens, and a switch that stops it at once.

Last updated

Four properties of every agent

  1. Its own identity. Substrait issues each agent a scoped, revocable identity. An agent never uses its owner’s standing credentials.
  2. Narrower than its owner. An agent’s authority is the smaller of its owner’s groups and what the agent was granted. Everything else is denied.
  3. A record of whom it acted for. Every action reads “this agent, acting for this person”. One accountable person stands behind every action.
  4. A switch on the identity. Revoking the identity stops the agent at once.

Checked as it acts

An app’s reach is known before it goes live; an agent’s is decided at runtime. Substrait therefore checks each action an agent takes, at the moment it takes it. That matters because content an agent reads, such as an email or a supplier’s document, can try to redirect it.

Who may operate agents

Operating an agent is a capability your admins grant to a person. The agent itself holds no seat.

How agents ask people

When an agent needs a human decision, it asks through Substrait’s communications centre. The right person answers from the channel they chose.

Related

Start with a result, not a rollout.

Book a demo and bring one real use case. We will show you how a tool your people build goes live on Substrait, with its data approved and its security check passed.