Agents as productivity leverage (24 tested PR's to production during naptimes)

Agents, and especially great harnesses, like Cursor, have taken the drudgery out of coding. I like solving problems, and the level of abstraction for problems does not matter. Some programmers struggle with the change, since a big part of their identity has been chiseling out the code.

At the same time, I'm grateful for the fact that I can code, so I can appreciate the tools we now have. It's not unlike music - you can appreciate it more, if you understand the complexity and skills required to produce it.

Kalle Tolonen
8.9.2026
Last updated on 9.9.2026

Things shipped by chatting with a computer

The most critical thing is that we've implemented a safe rollback mechanism, extensive test suite and as such, a robust CI-workflow to make these things possible for a production workload, and thus be able to push constant updates to prod.

Some of the work we've done during naptimes

Great UX improvements for the article writing experience
Scheduled social media posting for external distribution of blog posts
Implemented SSO
Added scheduled article publishing
Added locally run web speech transcription for dictated articles
Added WebP image pipeline
A long list of todo-chores from years of backlog
Enhanced an automated, secure, gated CI-pipeline from a VPS-hosted agentic harness to production via code repository
Security hardening

Key learnings

Modern agent harnesses don't require the user to be ultra specific. Over-speccing can actually hurt the outcome. It's quite feasible to have rapid deployments reliably (just ask Amazon). I have a legacy stack running Django, which is not that fast and does come with batteries included, and still most of the total deploy times are under a minute with slow Github runners, while running an extensive test suite, security scans and file transfers to a tiny VPS. One of the key things in any sort of optimization is optimizing only after you've gotten rid of ballast and automating only when you're sure about what you're doing. These things still require human taste and input.

You can and should be ambitious. DHH is building an agentic Linux distro that will eat Windows and MacOs away, since Linux has one key advantage. It's defined by text in files instead of fancy and slow GUI's. It's perfect for agents.

DIY

AI will build software for your business. Why use expensive SaaS, when you can make the software you need yourself or hire a cheap ass consultant to do it for you for pennies on a dime? If you outsource, pay for results, not time!