When not to use an agent
Expected question
"Everyone wants an agent. When do you refuse — and ship a workflow, a classifier, or a RAG answer instead?"
Variant forms
- "The VP said 'just make it an agent.' What do you say in the room?"
- "Workflow vs agent vs LLM-as-classifier — pick one and what would reverse you."
- "Our agent is a 40-step tool loop around a deterministic EDI map. What's wrong?"
- "When is RAG + a button enough?"
- "How do you stop agent-washing on the roadmap?"
- "What's the cheapest thing that could work for this customer in 90 days?"
- "Isn't a bigger model with tools always better?"
- "When would you take an agent out of production?"
The question, as it might actually be asked
"We have a LangGraph. Should this be an agent?"
Executive summary
30-second thesis
I'd use an agent only when the next step is bounded ambiguity and tools can change the world. If the path is a known state machine, a schema, or a retrieve-then-cite answer, I'd ship that. An agent is a cost and incident surface, not a compliment.
2-minute answer
Three tests, out loud. (1) Is the control flow known? If ops can draw it on a whiteboard without "it depends on the model," that's a workflow — durable, testable, cheap. (2) Is the output a label or a number? Calibrated classifier or ranker, not a tool loop. (3) Is the user asking a question over docs? RAG with access-before-rank and a decline. Agents start when you cannot write the branch table and a tool might refund, file, or page.
What I'd refuse: wrapping a deterministic map in a ReAct loop so the roadmap can say "agentic." That's 2026 agent-washing. Also refuse an agent whose only tool is search — that's RAG with extra latency.
Reversal: if the workflow's exception rate explodes and humans already do unstructured triage, then a bounded agent with gateway/HITL on the write. Promotion through evals (34), not a launch-day personality.
Clock in an HM screen: 90 seconds. In a design round: name the default in the thesis, don't tour patterns.
The framework
| Shape | Use when | Don't use when |
|---|---|---|
| Workflow / state machine | Known steps, SLAs, retries, humans at named pauses | The next action is genuinely open-ended |
| Classifier / ranker / rules | Label, score, route — you can define a metric | The user needs composed language and tools |
| RAG + decline | Questions over an authorized corpus | Irreversible writes, or the corpus isn't ACL'd |
| Agent + gateway | Bounded ambiguity + tools; eval can fail a promote | The branch table already exists and is stable |
| Human only | Prohibited / high-risk with no oversight you can staff | You're hiding a staffing problem behind a model |
What I'd say to the VP
If we can draw the steps, we should run the steps. I'll put an LLM on the exception pile and on the draft, not on the happy path. If we still want an agent, here's the HITL queue and the $ cap — not a demo.
Kill criteria (take the agent out)
- Exception rate is actually a missing field in the source system — fix the system of record.
- Reviewers rubber-stamp because volume beat staffing (34).
- Cost per successful task > a junior doing the workflow.
- Eval slice regresses and you cannot pin a digest (35).
What I'd ask them
- Can ops draw the happy path without the model?
- Which tool is irreversible on day one?
- What's the 90-day metric — handle time, error rate, or a slide that says "agent"?
What's expected at each level
- Mid-level: Proposes an agent because the prompt said agent.
- Senior: Mentions workflows vs agents; still defaults to a graph.
- Staff+: Picks workflow/RAG/classifier with reversal evidence.
- Principal: Refuses agent-washing in the room; names who inherits the default; kill criteria to remove an agent. Clock: 90s in HM, thesis in design. Fail-closed: no write path without gateway/HITL just because "it's only an agent."
Follow-up questions to expect
- "Isn't the model cheap enough that we should always agent?" Tokens aren't the incident. A wrong refund is.
- "What about a supervisor agent over workflows?" Fine as routing. The leaves stay workflows until ambiguity is real.
- "Customer demanded agents in the RFP." Ship a walking skeleton that is a workflow with an LLM draft. Don't lie on the RFP; don't light money on fire to match a noun.