Agents

An agent is software that can choose how to pursue an objective within boundaries you set. It is not simply a prompt with access to a model.

CopyAgent = objective + context + state + tools + execution + policy + evaluation

Objective

The objective describes the job and the definition of a useful result. It should be specific enough to guide decisions and clear enough to reveal when the agent must stop or hand work to a person.

Context and state

Context gives the run the facts it may use now. State carries the information needed to continue the current execution. Durable memory is for facts that should survive a run and remain scoped to the right agent, entity, and workspace.

Tools

Tools are structured capabilities an agent can call. They should have explicit inputs, outputs, permissions, and effects. A read tool retrieves context. A write tool changes something outside the agent and therefore needs an idempotency and approval story.

Execution and policy

The runtime enforces practical boundaries such as time, spend, concurrency, permissions, and sandboxing. Policy decides whether a specific action is allowed in the present context. These controls should not rely on the model remembering an instruction.

Evaluation

An agent improves through evidence. Test versions against real cases, compare the result and tool use with the expected behaviour, then promote only the version that meets the requirement.

Agents are appropriate when the route to an outcome requires judgement. For an explicit business process, use a workflow.