Be Unreasonable

Author: Thorsten Ball (works on Amp, the coding agent that began at Sourcegraph) URL: https://x.com/thorstenball/status/2113647306998264109 One-line: You are no longer working with humans, so stop asking agents for what you’d politely ask a coworker: demand exhaustive proof, ten variations, a bug tested to destruction, then walk away.

Posted on X on October 23, 2026 as a screenshot of text, with the caption “Don’t forget: you’re no longer working with humans” (1,355 likes when saved). See the tweet page. The post is short and is the author’s own, so it is transcribed in full below from the image.

Key claims

• Calibrate to the agent, not the coworker. Requests that would be “unreasonable, or over the top, or downright rude” to a human colleague are now normal to make of an agent.
• Demand proof, not reassurance. Ask for end-to-end retesting, an 8x8 test matrix with evidence (screenshots, demo videos, logs) in every cell, and a demo video, because the agent has “the patience and nothing else planned for the day.”
• Replace back-and-forth with breadth. Instead of iterating on a design (“maybe make it larger?”), ask for ten variations in different styles, presented on an interactive site where you can combine them.
• Agents put Big League practice within reach. A hard distributed-systems race condition can get the full treatment: reproduce, fuzz, instrument production, build a custom harness from the collected data, gate the fix behind a flag, roll it out in stages, and recheck the data.
• His own example: to build a feature for Amp’s experimental Gitlab support without being a Gitlab user, he handed the agent an SSH key to a throwaway box and told it to set up a Gitlab instance, reproduce the user’s setup, build the feature and prove it end to end. Then he walked away.

Full text

Transcribed verbatim from the screenshot (typos and doubled words as in the original).

Be unreasonable.

You have to constantly remind yourself that you’re no longer working with humans. Things that would’ve seemed unreasonable, or over the top, or downright rude — well, you can ask for all of it now.

“I want you to test this again. End to end. Again. Give me irrefutable proof that this now works.”

You’d be rude to ask that of your coworker, wouldn’t you? Not in Germany or Eastern Europe, of course, where this would be considered a normal code review comment on a Monday morning (if it’s the reviewee’s birthday). But everywhere else? Unreasonable.

Now it’s normal. You can ask that now. And you can go even further. Add this:

“Spare no token. Record a demo video. I want to go to sleep tonight knowing that you worked all day making sure this works. Give me an 8x8 test matrix of how to make sure this thing now works and then run through the whole matrix. I want you to put proof into each cell — screenshots, demo videos, logs, whatever.”

Make ‘em work, for they have the patience and nothing else planned for the day.

You’re not sure about the design of something? And you’re iterating with the agent, going back and forth, saying stuff like “hmm, not sure, maybe make it larger?” Don’t! Be unreasonable! Ask the things you’d never ask your designer colleague. Like this:

“Give me 10 variations of this UI. That’s right: ten. Explore different design approaches, different color schemes. Show me how they’d fix this at Apple, at Google, at Teenage Engineering. What would Rasmus Andersson do? Show me. Give me an interactive website on which I can look at all the the variations. Let me combine them.”

Be unreasonable! Ask for more things you ever asked for.

You can now do and build things that previously only the Big League could do.

Have a nasty distributed systems bug? A race condition that somehow involves multiple threads and how they’re scheduled and a lock? But you can’t figure out? Be unreasonable! Like this:

“Fix this bug! Find the race condition and then test this to hell and back. 8 times. At least. Try and break it. Use the debugger, use logs, fuzz if you want. Then add logging and data collection to the production system. Roll that out. Then, once it’s out get the data back. Then use that data to build a custom test harness for this bug and its fix so you can confirm it actually fixes it. Then I want you to wrap this rewrite of the function in a condition so that only I will run it in prod. Then roll it out. Then let me test it, then collect the data again. Then let my team try it. Collect the data. Then check whether the data contains new insights. Any crashes? No? Then let’s go.”

Be unreasonable!

Yesterday I was trying to add a new feature to the experimental Gitlab support I’ve added to Amp, but I’m not a big Gitlab user. I don’t have an account, really. You know what I did? I was unreasonable.

I told the agent: “Here’s an SSH key to a throwaway box. I want you to SSH into that, set up a Gitlab instance, make sure it works, then reproduce the setup this user who asked for the feature has, make sure our integration works end to end, then add this feature and test that end to end. I want irrefutable proof that you made this feature work. Dazzle me!”

And then I walked away and did something else. Completely unreasonable.

How it connects

• Thorsten Ball: author; works on the Amp coding agent, which began at Sourcegraph.
• AI Coding / AI Agents: a practitioner’s rule of thumb for delegating to coding agents, where the scarce input is the human’s attention, not the agent’s effort.