Insight · ·

AI agent or deterministic workflow: who chooses the next step?

Use a fixed workflow when the path is known. Introduce an agent only where choosing the next action is part of the problem.

AI agent or deterministic workflow: who chooses the next step?

Keep known rules out of the model's discretion

Choose a deterministic workflow when valid inputs and business rules determine the next step. Consider an agent when the system must discover which action is useful from incomplete information. Calling a fixed process an agent does not improve the process.

Anthropic draws this architectural distinction between predefined workflow paths and agents that direct their own tool use. The distinction is useful even when neither implementation uses Anthropic's products. Anthropic on workflows and agents

Deterministic orchestration does not mean every component is non-AI. A model can classify an incoming message inside a fixed workflow. The application can still validate the classification, apply policy and route uncertain cases for review. The model need not control the entire process.

Find the uncertainty that actually needs reasoning

An illustrative invoice process has predictable stages: receive a document, extract fields, validate the supplier and request approval where required. The text extraction may be uncertain; the rule that approval is required should not be.

A hypothetical investigation into an unusual service complaint is different. Useful evidence might live in an order record, a delivery update or a support exchange. The next search can depend on what the previous search revealed. That is a credible candidate for bounded agentic investigation.

Do not use an agent to conceal an undocumented process. If staff disagree about the approval rule, model discretion does not settle the policy. Resolve the ownership question first, or the automation will convert disagreement into inconsistent outcomes.

Define authority separately from intelligence

A system that can identify a likely remedy does not automatically have permission to execute it. Keep recommendation, authorisation and execution distinct.

An investigating agent may gather permitted evidence and propose a resolution. A deterministic service can check account ownership, permitted amounts and required approvals before performing the operation. The agent's fluent explanation is not the authority for the write.

The least-privilege agent guide covers narrow tools and scoped credentials. Those controls remain necessary even if the model is excellent at the task. Intelligence and access are different design decisions.

Define stopping conditions as well. An agent should be able to report missing evidence, hand off an unresolved case and stop without treating an arbitrary guess as completion.

Compare accepted outcomes, not impressive traces

Evaluate both approaches on representative tasks, including ambiguity, missing information and an unavailable dependency. Count whether the business outcome is acceptable, how much intervention it requires and whether the system explains why it stopped.

A long reasoning trace is not evidence of a better result. A fixed workflow that reliably escalates an unsupported case may be more useful than an agent that produces a confident but ungrounded resolution.

Costs include investigation steps, retries, review and operation after failure. Latency also matters when a person is waiting for the result. Avoid assuming that more autonomy creates more value; it may simply relocate work from implementation into supervision.

Draw the boundary this week

Take a real process and mark each step as a known rule, an uncertain interpretation or an open-ended investigation. Keep the known rules deterministic. Test the uncertain interpretation in isolation before granting control over later actions.

Prototype the investigation with read-only tools, a clear stop condition and a human-reviewed result. Expand authority only after the evaluation supports a specific need and the recovery path exists.

Clodron's AI automation work should use the smallest amount of model discretion that solves the problem. The right outcome may be an agent, a fixed workflow with an AI component, or no AI at all.