Agent Primitives
Agent Primitives are the building blocks an enterprise needs before it can trust an agent to act on its systems and data. They come from what it took to build and run such agents. Each one assumes that agents are multiplayer, that an agent acts on its own while a person approves the consequential steps, and that the model runs inside a strict boundary. The boundary prevents unauthorized actions and keeps data from leaking.
Each primitive below covers one concern every enterprise agent has to solve, from how an agent is defined to what it may remember. Its full definition sets out the requirements a system has to meet, what still holds with the model compromised, and the systems commonly used in its place that don't qualify.
- Agent Definition
An agent is two objects with two lifetimes. A durable declaration states what the agent is. An ephemeral instance backs one session.
- Safe Tools
A tool is a named slice of a capability. The operations it permits, and the arguments they may carry, are fixed before the agent runs. The agent cannot widen that list without human approval.
- Identity and Credentials
An agent acts as itself, or on a person's behalf. It gets the use of a secret without being given the secret.
- Authorization
Every action resolves to one subject, one permission, and one resource. The action proceeds only on an explicit yes for that triple, at the instant of the call.
- Channels and Continuity
The conversation is the session. It is also the surface where authority over its work is checked. Each message and each decision is a claim, and the system checks that claim before it takes effect.
- Memory and Knowledge
One store holds both what an agent has learned and the record of what it did. It is an authorization surface and an evidence record first, and a retrieval system second.