Start with the shape of the work
Automation choices should follow the process. A stable sequence with explicit inputs and rules is usually best expressed as deterministic workflow logic. Language understanding or classification can be inserted as a bounded AI step without turning the whole process into an agent.
Agentic behavior becomes more relevant when the system must choose among tools, adapt a plan after observing results, or handle common variation that would make a rule tree unmanageable.
Use the least autonomous design that solves the problem
Autonomy creates flexibility, but it also expands the system’s possible trajectories. More trajectories mean more evaluation, permission, tracing, and recovery work.
A strong design often combines deterministic orchestration with a bounded agent inside one part of the workflow. The outer process controls state, timing, and approval; the agent handles the variable reasoning task.
- Rules for known calculations and policy boundaries
- Models for extraction, classification, or recommendation
- Agents for adaptive multi-step tool use
- People for accountable consequential decisions
Authority must be designed explicitly
A tool-using agent needs more than API access. Each tool should expose only the actions and data required for the task. Financial, legal, security, customer, and irreversible actions need appropriate approval and idempotency controls.
Human-in-the-loop design is not a single confirmation screen. It defines when review is needed, what evidence the reviewer receives, how an override is recorded, and what happens when no reviewer responds.
Evaluate trajectories, not only final text
An agent can reach a plausible final response through an unsafe or wasteful path. Evaluation therefore needs to inspect tool choice, arguments, ordering, retries, policy compliance, and the evidence used to stop.
Production tracing should let operators reconstruct the path without exposing sensitive reasoning or data unnecessarily. The observable artifact is the sequence of state and action—not a promise that the agent “thought correctly.”
The practical boundary
Choose deterministic workflow when the path and rules are known. Add AI at bounded steps when perception or language is variable. Use an agent when adaptive planning across tools materially improves task completion. Keep accountable people in control of consequences the system should not own.