Agent or workflow? Choosing the autonomy level

A workflow is a system where your code decides the sequence of steps and models fill in some of them. An agent is a system where the model decides the next step and which tools to use. Between the two sits a spectrum of patterns: a single model call, prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimiser and the autonomous agent. Each step towards autonomy buys flexibility and pays for it in predictability, auditability, latency, cost and the size of the blast radius when something goes wrong. In most enterprise processes the steps are already known, the rules are already written down, and the genuinely uncertain part is one judgement: classify this, extract that, summarise the rest. That is a deterministic workflow with one model call. Where you do use agentic steps, the model proposes and deterministic code enforces.

Key takeaways:

• The operational test: if the model can change what happens next, that part is agentic. If code decides, it is a workflow.
• Start with the simplest design that meets the acceptance criteria and add autonomy only when evals show it helps.
• The determinism boundary is the most important line in your architecture diagram. Draw it explicitly.
• Autonomy is set per action, not per system. Read tools and write tools sit at different levels.
• Hybrids are normal: a workflow with an agentic step, or an agent whose proposals flow into a deterministic executor.