An agent system combines a model, instructions, tools, state, observations, and a runtime that controls execution. The model proposes a next step. The runtime validates the proposal, checks permissions and budgets, invokes a permitted tool, and records the result. A final response or a stopping condition ends the loop.
State records what is known and what has happened. Memory may preserve selected information across tasks, but it requires ownership and retention rules. Tools expose narrow capabilities such as reading a ticket; they should not automatically expose every operation the backend can perform.
The environment can fail or change while a task runs. A tool timeout, revoked permission, or user cancellation must produce an explicit state. A convincing final sentence cannot substitute for knowing whether an action actually completed.
Exercise: label model, runtime, tool, state, and observation in a three-step support workflow. Add a step limit and a condition requiring human review.
Check: you can identify which component enforces permissions and which component stores the record of an executed action.