THE SHORT ANSWER
Use a fixed workflow when the steps are known. Add AI inside bounded steps when interpretation helps. Use an agentic workflow when the model must choose among constrained next steps. Consider higher autonomy only when variation demands it and observation, permissions, cost limits and escalation remain strong.
Four levels of model-directed choice
| Pattern | Model discretion | Useful when | Control burden |
|---|---|---|---|
| Fixed workflow | None or one bounded inference | The path is stable | Lowest |
| AI-assisted workflow | Interprets within named stages | Inputs vary but process does not | Validate model outputs |
| Agentic workflow | Chooses among allowed branches and tools | Evidence changes the next step | Observe paths, tools and limits |
| Higher-autonomy agent | Plans and adapts across a wider action space | The useful path cannot be specified well | Highest permissions, evaluation and oversight burden |
Evidence & context: Anthropic
Choose by uncertainty and consequence
Variability argues for more flexible interpretation. Consequence argues for tighter control. A variable but low-impact research task may tolerate wider exploration. A predictable financial update should remain narrow even if a model could perform it.
Volume, latency and cost also matter. Fixed paths are easier to cache, batch and test. Agentic paths can add model calls, tool calls and retries whose value must be measured at the completed-task level.
Combine patterns deliberately
A marketing workflow might use deterministic campaign limits, a model to classify a brief, an agent to research uncertain questions and a human to approve publication. The system is stronger because autonomy changes by stage, not because one label governs everything.
Start at the lowest level that can plausibly meet acceptance criteria. Move upward only after representative cases show where fixed paths fail.
Make autonomy a reviewable decision
- Which decisions cannot be specified in advance?
- What new actions become available at this level?
- Can each action be reversed or independently verified?
- How does the system behave when information conflicts?
- What is the maximum time, cost and number of steps?
- Where does a person retain authority?
Evidence & context: NIST
Sources & further reading
- Building effective agents
Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
- Using tools
OpenAI Developers. Official documentation showing how models can be given built-in, function and remote tools. Checked 13 September 2026; product-specific tool names and availability can change.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗