Agent

An agent is a scoped unit inside an agentic system. It can adapt a plan as context changes, but only within explicit input, tool, action, budget, escalation, and stop boundaries. Multi-agent decomposition is used only when specialization, parallelism, or context isolation earns its added operational cost.

What an agent has

An agent's exact shape follows the work. It typically has a scoped view of approved data, only the tools and skills justified by its mandate, and explicit input and output contracts. State and handoffs use the simplest controlled mechanism the architecture requires; direct calls, workflow state, durable events, queues, or another justified pattern may fit different boundaries. A2A is used only when an independently operated agent-system interoperability boundary justifies it.

What an agent is not

An agent is not a chatbot by default. When an agent is used in an operational workflow, its structured inputs, outputs, tool calls, latency, evaluation results, and material actions should be observable under the deployment's agreed controls.

An agent is not assumed to be unsupervised. Oversight follows risk, impact, uncertainty, and reversibility: a low-impact action may run inside an agreed permission and rollback boundary, while consequential or ambiguous work routes to a named expert review path. The customer retains decision accountability and defines that line in the Decision Blueprint.

How many agents per system

More agents is not better. The right count is the smallest number that produces a measurable quality or operating benefit; some workflows should remain deterministic or use a single agent.