Build agents

An agent combines an objective, scoped context, state, tools, execution limits, policy, and evaluation. It is appropriate when the route to a result needs judgement; use a workflow when the path should be prescribed.

The platform agent layer is built around specialist execution, capability discovery, structured tools, explicit contracts, policy checks, and auditable runs. Publicly, you build against those capabilities rather than internal service families.

Start narrow

Give an agent one useful job, read-only tools, a limited entity scope, and a clear stop condition. Run it against cases you know. Only then add an action and, if needed, an approval.

See Tools and context and Runs and versions.