Deterministic workflow
The steps and branch conditions can be specified in ordinary code. Inputs may vary, but the orchestration does not need open-ended judgment.
LESSON 01 · FOUNDATION
Agent or Workflow?
Before adding planning, memory, tools, or autonomy, determine whether the business problem benefits from agency at all.
Start with the structure of the work.
Teams often call any LLM-backed automation an “agent.” That label hides the most important design decision: who chooses the next action?
If the sequence of steps is already known, encode the sequence. If language interpretation is uncertain but the process is fixed, use an LLM inside a workflow. Reserve an agent for problems where the system must repeatedly observe state, choose among possible actions, inspect the result, and decide what to do next.
D→V RULE
Use the least agency that can reliably change the target outcome. More autonomy increases the action space you must evaluate, observe, secure, and recover.
LLM workflow
Language or unstructured evidence requires model judgment, while the surrounding sequence, tools, and stopping conditions remain explicitly orchestrated.
Agent
The system pursues a bounded goal through a feedback loop. It decides which permitted action to take next based on changing state.
01Observe
02Interpret
03Choose
04Act
05Verify
06Continue / stop
INTERACTIVE LAB · CLASSIFY THE SYSTEM
Three problems. Three different control structures.
Choose the smallest architecture that fits each problem. Your score rewards the reasoning pattern, not maximum autonomy.
SCENARIO 01 · REFUND CORRECTION
Duplicate webhook delivery
A payment provider may deliver the same refund event more than once. The business requirement is that the ledger records one refund and repeated delivery returns the same result.
PATHknown
AMBIGUITYlow
SIDE EFFECTfinancial
SCENARIO 02 · POLICY INTAKE
Classify and draft a case
An operations team receives free-text customer messages. The system extracts intent and evidence, checks a fixed set of records, drafts a response, and always routes the draft to a human reviewer.
PATHfixed
AMBIGUITYlanguage
SIDE EFFECTdraft only
SCENARIO 03 · FREIGHT EXCEPTION
Investigate a moving operational problem
A shipment is late. The system may inspect shipment events, weather, terminal status, customer SLA, and available recovery options. New evidence can invalidate earlier hypotheses, so the next useful tool call depends on what the system just learned. Any external commitment requires human approval.
PATHdynamic
AMBIGUITYhigh
SIDE EFFECTgated
DESIGN TEST · AGENT SUITABILITY
Score the need for agency.
This is a design aid, not an automatic approval gate. High scores mean that dynamic action selection may be useful; they do not mean the system should receive broad authority.
LESSON ARTIFACT
Write the Agent Suitability Record.
The artifact forces a team to justify agency before implementation. It can later feed the Agent Boundary and Agency Budget lessons.
Sample · freight exception investigation
- PROBLEM
- Investigate late shipments before a dispatcher commits to a recovery action.
- OWNER
- Dispatch lead.
- CONTROL
- Agent for investigation; human approval for external commitments.
- WHY
- The useful next evidence source depends on what was learned from previous calls. A fixed tool sequence would either over-query or miss relevant branches.
- PERMITTED
- Read shipment events, weather, terminal status, customer SLA, and available recovery options. Produce recommendations. Do not change bookings or communicate commitments.
- STOP
- Escalate when evidence conflicts materially, required records are unavailable, confidence is insufficient, or the proposed action would create an external or financial commitment.