AI System Design · Designing Agentic Systems · Lesson 2 of 2
When NOT to use an agent
The most senior thing you can say about agents is when not to build one. Agents trade predictability for flexibility — and interviewers plant tasks that look agentic but shouldn't be.
Reach for a workflow or single call when...
- The steps are known and fixed — a hardcoded chain is cheaper, faster, and more testable.
- You need predictable latency and cost — agent loops are variable by nature.
- The task is high-stakes with low error tolerance and doesn't need autonomy — constrain it.
- You can't yet evaluate the trajectory — shipping an unmeasurable agent is how you get silent failures.
The ladder, againSingle call → workflow → single agent → multi-agent. Each rung adds capability and subtracts control, latency headroom, and reliability. Climb only when the rung below genuinely can't do the job.
Agent-washing a workflow"Design an agent to send an invoice email after payment" is a trap. Payment→email is a fixed two-step workflow. Adding an autonomous agent there is pure over-engineering — say so and propose the workflow.
A crisp rule: use an agent only when the path is genuinely dynamic (the model must decide which tools and how many steps at runtime) and you can measure whether it succeeded. Otherwise, constrain it.
◆ Lock it in
- Use a workflow/single call when steps are fixed, latency/cost must be predictable, or you can't eval the path.
- Every rung up the ladder subtracts control and reliability.
- Calling out an over-engineered agent is a strong senior signal.
Feynman drill — say it out loudGive one task that truly needs an agent and one that looks agentic but is really a fixed workflow.
Step 1 rate your confidence · Step 2 pick your answer
Which task is best served by a fixed workflow rather than an autonomous agent?
Step 1 — how sure are you?