01
Autonomy is a risk decision
Agents are attractive when a task requires interpretation, tool selection, or flexible sequencing. They are a poor default when the process has stable rules, consequential side effects, strict timing, or a narrow set of valid states.
The useful question is not whether a process can use an agent. It is which decisions benefit from flexible reasoning and which must remain explicit.
02
Prefer deterministic orchestration when
A state machine or workflow engine provides stronger control when the happy path and important exceptions can be enumerated. It supports replay, testing, service-level commitments, and operational ownership.
- Inputs and transitions are well defined.
- The same evidence should produce the same result.
- Actions carry financial, legal, security, or customer consequences.
- Operators need predictable recovery and replay.
03
Use bounded intelligence inside the workflow
AI can interpret unstructured input, draft a response, rank candidates, or explain a rule while deterministic orchestration controls when tools run and which actions require approval.
This hybrid design often delivers most of the value attributed to agents with clearer evaluation and a smaller failure surface.
Decision guide
Questions to carry into the work
- Map states, transitions, exceptions, and side effects.
- Identify steps where ambiguity prevents reliable rules.
- Bound model inputs, outputs, tools, time, and cost.
- Require approval before consequential actions.
- Design timeout, retry, fallback, and replay behavior.