Matching control flow to the task
Lesson 5 of 5 in State Machines vs LLM Loops: Who Owns Control Flow?.
Time to make the spectrum operational. The deciding questions are always the same three: can you enumerate the steps in advance? — what does a wrong action cost? — who has to be able to inspect the run afterwards? High enumerability pushes toward code; high wrong-action cost pushes decisions behind coded gates; audit obligations push state into the open. Here are five recurring task shapes and where they land.
| Task shape | Control-flow style | Why | Watch out for |
|---|---|---|---|
Regulated claim / KYC processing | Fully-coded state machine, model filling slots inside states | Steps are legally fixed; replay and audit trails are non-negotiable; every transition may need a named approver | Rigidity meets creative reality — build an explicit exception state or watch edge cases get misfiled |
Customer support triage & resolution | Hybrid — coded skeleton, model judgment in the nodes | Escalation paths and refund limits must be guaranteed; diagnosing what the customer means takes reading | Model verdicts must cross a schema; low confidence routes to a human, never to a guess |
Open-ended research & synthesis | Free-running LLM loop | The useful next step depends on what the last source revealed — the step space cannot be enumerated up front | Cost and latency runaway: budget caps, step limits, and a kill switch come from the runtime, not the prompt |
Nightly document-extraction pipeline | State machine / plain workflow — possibly no agent at all | Same steps every night on predictable inputs; a model may do the extraction inside one step, but nothing needs to decide anything | Autonomy nobody asked for — adding a loop here buys failure modes, not capability (see the when-not-to-build-an-agent crosswalk) |
Coding & debugging tasks | LLM loop with a verification harness | Compilers and test suites give the loop cheap, honest grounding every iteration — the rare setting where free-running autonomy is well-policed by the environment | The environment checks correctness, not safety: sandboxing and explicit stopping conditions still come from you |
Interactive sorting exercise: Sort each task into the control-flow style you would reach for first.
Tool: Architecture Advisor — Take the spectrum for a spin: describe a task and defend a control-flow choice against the advisor’s follow-up constraints.