The anti-checklist: signals you do not need an agent
Lesson 3 of 4 in When NOT to Build an Agent.
Checklists usually tell you when to do something. This one is inverted: any single signal below is strong evidence you should not build an agent — no weighing, no scoring. One hit is enough to send you back down the ladder to a workflow or a script.
Run every “we need an agent” request through these five before anyone opens an IDE.
Signal 1 · The steps are fixed
If you can draw the complete flowchart — every branch, in advance — you have just designed the workflow. Building an agent now means paying the model, on every single run, to re-derive a diagram you already own. The whiteboard test is brutal and reliable: teams asked to actually draw the “unpredictable” process usually finish the drawing.
Signal 2 · There is only one tool
An agent’s job is choosing — which tool, with what arguments, in what order. A loop that only ever calls one tool has nothing to choose between; it is a retry loop wearing a lanyard. Wrap the single call in ordinary code with ordinary error handling and keep the determinism.
Signal 3 · The output must be deterministic
Payroll, billing, tax computation, regulatory reports: same input must produce same output, and someone must be able to certify why. A probabilistic system cannot make that promise — not with temperature zero, not with better prompts. Non-determinism is the price of model judgment; if the requirement forbids the price, the requirement has answered the question.
Signal 4 · You need sub-second latency
An agent thinks in seconds to minutes, and the loop’s length is not yours to set. Autocomplete, fraud checks in the payment path, anything a user watches a spinner for: the model can still help — offline. Precompute with models, serve from code. Never put a loop in a hot path.
Signal 5 · Zero tolerance for error
If a single wrong action is unacceptable and cannot be caught by a gate before it takes effect, a system with an irreducible failure rate does not belong in the action path. Note the two-part test: zero tolerance alone might still permit an agent whose every action is gated — that is the “agent with gates” outcome in the next lesson. Zero tolerance plus ungateable actions is a hard no.
Interactive sorting exercise: A requirement lands on your desk. Sort each one: does it point to something simpler, to an agent, or to an agent whose every consequential action is gated?
Key terms: anti-checklist, tool, non-determinism, approval gate, workflow
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.