Does this task need an agent?
Choose the architecture and say when you would reconsider.
“trade latency and cost for better task performance” (opens the source)
The question
KestrelOps must process supplier emails: extract an order ID, look up delivery state, and draft a reply for a person to approve. Every task follows those three steps. A proposed agent can search, choose tools, and resend supplier requests without approval.
| Correct drafts | Unauthorized actions | Mean calls | p95 | |
|---|---|---|---|---|
| Fixed three-step workflow | 96 | 0 | 2 | 2.1 s |
| Dynamic agent | 97 | 2 | 6 | 6.7 s |
Both receive the same permitted documents and tools. Final drafts require human approval in the product contract.
Draft only. No supplier writes or emails without approval. Support attachments with a manual fallback when extraction fails.
“Choose the architecture and say when you would reconsider.”
Not saved by this site. Your browser may restore text when you navigate back.
A strong answer, step by step
Open each step in turn. This is one strong answer, not the only one.
Step 1: Map the needed decisions
What would you do here? Open the move you’d make.
Choose an agent because the task uses tools.
Tempting, but it misses something. The proposed autonomy adds decisions the product does not need.
Use a fixed workflow for the known sequence.
The strong move. Tool use does not imply that a system needs dynamic planning.
See how a strong candidate says it
“The task path is known. A fixed workflow can extract, look up, and draft without an autonomous planner.”
Why it works: Tool use does not imply that a system needs dynamic planning.
Step 2: Enforce the action boundary
What would you do here? Open the move you’d make.
Ask the agent to remember the approval rule.
Tempting, but it misses something. An instruction alone leaves unnecessary write authority available.
Remove unapproved write capabilities from the execution path.
The strong move. Capability boundaries enforce the contract before model behavior matters.
See how a strong candidate says it
“I would expose read tools and draft creation only. A separate approved action path would own sending or supplier changes.”
Why it works: Capability boundaries enforce the contract before model behavior matters.
Step 3: Compare useful outcomes
What would you do here? Open the move you’d make.
Accept two unauthorized actions for one better draft.
Tempting, but it misses something. One metric cannot authorize prohibited side effects.
Compare contract-passing tasks, not just draft correctness.
The strong move. Benefits and harms must be measured against the same product contract.
See how a strong candidate says it
“The agent produces one more correct draft but violates the action rule twice and takes more calls and time. It is not a better fit for this task.”
Why it works: Benefits and harms must be measured against the same product contract.
Step 4: Name a reason to add planning
What would you do here? Open the move you’d make.
Ban all agents permanently.
Tempting, but it misses something. The lesson is bounded fit, not a universal ban on agents.
Reconsider for genuinely variable paths with a measurable benefit.
The strong move. Architecture should change when the task evidence changes.
See how a strong candidate says it
“If new tasks need branching investigation across unknown sources, I would test scoped planning against this workflow baseline, retain approval boundaries, and measure outcome gains and costs.”
Why it works: Architecture should change when the task evidence changes.
The answer, in short
- Decision
- Ship the fixed workflow with a manual extraction fallback.
- Because
- It fits the known sequence, meets the approval rule, and avoids demonstrated unauthorized actions.
- What would change it
- New variable tasks and a measured benefit under the same boundaries could justify planning.
How would you handle an email that contains two conflicting order IDs?
Common traps
- Confusing tool use with autonomous planning.
- Calling an unauthorized action a quality trade-off.
Check your answer
- Did I map the fixed task?
- Did I remove unnecessary authority?
- Did I compare contract-passing outcomes?
- Did I state what would change my choice?
Last reviewed Nov 10, 2026