06Economics & tradeoffsDeep dive ②

The Agent Isn't Expensive. Your Task Selection Is.

Every agent call carries a fixed overhead cost — and for tasks below a certain threshold, that cost eats your savings whole.

Read4 min read
Topicseconomics · overhead · task-selection · delegation
TL;DR

Every agent call carries a fixed overhead: write prompt → wait → read output → iterate if needed. Tasks complex enough make that overhead worthwhile. Tasks simple enough don't. "Faster" isn't a property of the agent — it's a property of the task you give it.

I once asked an agent to write a thank-you email after a meeting. Four minutes to open a chat, frame the prompt with enough context, wait, read the output, adjust the tone, copy it out.

I could have typed that email myself in forty-five seconds.

The agent wasn't bad at it. I just handed it the wrong job.

01The Overhead Isn't Optional — It's Fixed

Every time you open a chat with an agent, there's a cost you can't avoid. Whether the task is large or small, whether the output is long or short, four things always happen:

Write a prompt with enough context — you have to describe what you need. Simpler tasks need less, but never nothing. Wait for the output — seconds to tens of seconds, depending on complexity. Read the output — you can't trust it blindly, even for simple things. Handle any needed revision — even simple tasks sometimes take a follow-up round.

In total, even the smallest agent call costs around two to four minutes of overhead. That's fixed — it doesn't shrink just because your task does.

So the real question becomes: does the savings exceed that two-to-four minute baseline?

02When Overhead Swallows the Savings

Agent usually net positive

Task takes more than 10–15 min by hand — overhead is a small fraction
Domain or format you lack fluency in — agent covers a real gap
Low cost-if-wrong: drafts, exploration, throwaway work

Agent often net negative

Task takes under 3–5 min by hand — overhead is most of the cost
Task you do naturally well: polite short replies, routine messages
You have to verify closely anyway — high stakes or context only you hold

Three dimensions to weigh: how long it would take you by hand, whether you have a natural edge, and what's the cost if it's wrong. You don't need all three to point the same way — but if they all lean toward the right column, that task probably isn't for the agent.

The problem is that plenty of tasks live in the grey. Writing a short product description: quick, familiar, but maybe you're not fluent in this market's tone. Each one is a small judgment call — and making that call takes less time than you think, if you have a frame for it.

03Signs You're Paying Overhead You Shouldn't

When overhead is eating your savings
You already know what the output should look likeif you've mentally composed it — you'll write it faster than you'll explain it to the agent
You're editing the output more than using itif every result needs significant adjustment, the "agent does it for you" part isn't as large as it felt
Your prompt took longer to write than the task would haveespecially when the context only lives in your head and takes a paragraph to transfer
You open the chat first and think of the task secondreaching for the agent by reflex, not by judgment — that's automation bias, not real optimization

The last one is the most dangerous: when using the agent becomes a reflex rather than a choice. Reflexes don't run cost-benefit analysis — they just run.

04Three Questions Before You Open the Chat

No elaborate process needed. These three questions, asked in ten seconds, are usually enough:

1
How long would I take doing this myself — honestly?

"Not imaginary time. The time your hands would actually need right now, including the thinking. If the answer is under five minutes — the overhead may not be worth it."

2
Does the agent have something I don't?

"Typing speed? You don't need that. Domain knowledge you lack? Yes. A tone or format you're not fluent in? Yes. If the agent doesn't have a real edge on this task — the overhead is pure cost."

3
Can I verify it quickly?

"If verifying the output costs as much effort as doing it yourself — keep it. For tasks where you'd read every word carefully after receiving them: doing it yourself is rarely much slower."

These three questions don't filter out the agent — they filter for the right tasks. The best fit is a task where all three answers lean toward "agent has a real edge." On those, it's clearly net positive.

There's a counterintuitive truth here: the most effective agent users aren't the ones who open the chat most often. They're the ones who know when to close the tab and do it themselves.

Overhead isn't the enemy. It's a natural filter that helps you use the agent where it actually belongs. The question is whether you let that filter run deliberately — or let reflex decide for you.

c
The author

Each story here wraps a lesson paid for in full.

craftagentsomeone building and learning at once

What are you building with agents? Want to trade notes, push back, or build something together — drop a line.

58pieces12clustersVI·ENbilingual

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.