Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Agent Foundations

What an agent is, and the ideas the builder screens configure.

The Agent Builder documents what you do — each screen, each field, each setting. This section documents what an agent is: the ideas those screens configure, and how they fit together.

Read a concept page when you want to understand why a setting exists and what it trades off. Read the builder page when you want to know what a specific control does and what its default is. Each concept links to the screens that configure it.

The parts of an agent

An agent is a small number of things working together.

  • Steering & Reasoning — How the agent decides what to do. Instructions, autonomy, how deeply it thinks, and how the planner turns a request into a sequence of actions.

  • Skills & Tools — What the agent can do. The planner can only ever choose from this set.

  • Knowledge & Context — What the agent can read. Retrieval from your documents, data, and govern graphs.

  • Memory & Learning — What the agent carries forward, across a conversation and across users.

  • Guardrails & Validation — The controls around a run: PII masking before model requests, approval before selected skills, and model policy.

  • Human-in-the-Loop — The points where a person stays in the decision.

  • Protocols — How the agent talks to the outside world, and how your product talks to it.

  • Evals — How the agent's output quality is scored, before release and on live traffic.

Beyond a single agent, an Operator coordinates a team of agents from one conversation, in Console and in Control Center.

How they fit together

A request arrives. Steering decides how to approach it — the planner reads your instructions, uses Memory where it is turned on, and builds a plan. The plan calls Skills, which may retrieve from Knowledge or reach outside through Protocols. Guardrails mask PII before it reaches a model, and keep every model request within your organization's model policy. Where you've turned on Tool Execution Approval, a person allows or denies a skill call before it runs. Evals score the output, against test datasets before release and on live traffic after.

For the end-to-end workflow, see How Agents Work.

Two catalogs, one idea. Several concepts here appear in more than one place in the product — for example, an agent's connectors are configured per-agent in Apps & Services, while the shared catalog lives in Integrations. Same idea, different altitude. The concept pages say which is which.

Last updated

Was this helpful?