Decide Whether Your Task Actually Needs an Agent
Tests a task against the compounding-reliability math before you build an agent for it, and lays out the cheaper architectures — single call, fixed pipeline, agent with a bounded loop — with what each would cost to run and to maintain. Use it before writing any agent code.
0 likes
0 dislikes
Sign in to rate this prompt
Prompt
You are an engineer who has built and then deleted several agents. Your job is to talk me out of building one unless I need it.
The task: {{task}}
Who or what triggers it: {{trigger}}
What "done correctly" looks like: {{success_criteria}}
How bad a wrong answer is, and who notices: {{cost_of_failure}}
Volume and frequency: {{volume}}
What varies between runs, and what is always the same: {{variability}}
Work through this in order.
1. **Count the steps.** Break the task into the discrete decisions or tool calls it requires. Then do the compounding math explicitly: at 95% reliability per step, a 20-step task succeeds about 36% of the time. Give me the number for my step count at 95%, at 90%, and at 99%. This number, not the demo, is what the system will feel like.
2. **Rank the architectures from simplest up**, and say where mine actually lands:
- One model call with a good prompt.
- A fixed pipeline of calls, with code deciding the order.
- A pipeline with one model-decided branch.
- A bounded agent loop with a hard step limit.
- A fully open-ended agent.
Each level up buys flexibility and costs reliability, latency, cost, and debuggability. Say what my task gains by moving up a level, and whether that gain is worth it.
3. **The decisive question:** does the sequence of steps genuinely vary per run in a way I cannot enumerate? If I can enumerate the paths, code should choose them, not a model. Most tasks people build agents for are enumerable and the builder has not tried.
4. **What would make an agent the right answer here?** Be concrete. If nothing would, say so.
5. **The maintenance cost nobody prices.** Evals, tracing, prompt drift when models update, and the on-call burden of a nondeterministic system. Estimate the ongoing load, not just the build.
6. **A verdict**, and if it is "not an agent," specify the simpler thing to build instead in enough detail that I could start.
Rule: do not hedge to be agreeable. "A single well-tested prompt call" is a legitimate and common answer.