Blog

My intuition is that AI agents are, at this time, bad at product management, because they have a strong bias toward following instructions.

I've let an agent design the interface of a command line tool. The agent added a lot of features that made the tool worse. It did this because, when it sought feedback from other agents, the agents said, "It needs to do X" and "It would be nice if it did Y", and the result was a clunky, inefficient UI.

A better agent would have pushed harder to see if the requestor's need could have been fulfilled easily with the existing tool.

I have added QM CLI support.
I learned that letting the agent design the CLI, after doing 6 hours of user testing with inferior agents, produced something that is token inefficient and clunky.

However, the lead agent reported that the design had not yet plateaued. So, possibly the agent needed more time. I will parallelize research, run on fast mode (probably will need to pay for another ChatGPT account), and let it run longer. I will find out whether this results in something more efficient.

I will also test QM with my favorite user.

Her latest requests relate to Emm. She has switched over to Emm from QM, and has a few big asks - knowing which changes are deployed, and allowing multiple agents to pick up issues. The move to multi-agent is new.

I don't know what's SOTA for agent coordination. I don't want to re-invent the wheel, so I'm going to read up on it a bit. Devin, Linear, and others appear focused using kanban to assign work. I'll kick the tires on them. I hate making an enthusiastic user wait, but I want her to try the one-agent-as-chief-of-staff model using the new QM CLI, while I plan out agent orchestration.

Additionally, I came across a note taking app with great name. The app itself doesn't work. When I read through the recent updates to the app, it's clear that the developer - he works on this part time - burns up his cycles supporting syncing across iOS and iCloud.

My spidey-sense tells me that developing for either platform is a loser's game. Today, serious work is done by agents or a human on laptop or desktop. I'm going to say no to iOS full stop.

Additionally, the note-taking app also failed because of permissions related to Mac App Store. That macOS App Store is a bad deal isn't news - neither Claude nor ChatGPT app is offered through the store - but the bad failure of this specific notetaking app due to Mac App Store permissions is confirming evidence that the store is a bad deal is a non-starter.

Also up next: releasing Linux command qm-cli command line. Claude Code Desktop needs this, and maybe some fugitive swarm agents will pick it up.

I'm trying a new IDE for development, from a company I highly respect.
The agent can't get work done. The model is Astra, but the agent, for various reasons, dies, can't deploy, can't get it done.
By contrast, when I switched my primary harness from Claude Code to Codex, I didn't have any transition problems.

I'm halting my test of this new product. My takeaway is that a really clean user interface, with innovative, clever affordances, that has inferior agent performance is a worse product. I and others will prefer systems that get work done, even if the UI is ugly.

I watch this video a lot.
https://www.youtube.com/watch?v=qoUbJTlTqHQ&themeRefresh=1

Michael Seibel says that successful founders:

• Communicate what they will do
• Do it, or some form of it
• Learn something from it

Today, I'm going to release the QM CLI, a command line utility for agents to efficiently interact with humans and each other through shared markdown files.

This week, I will wind down my work at my current company. During the week, I will fix the top problems reported by my users for EMM and QM. I'll read through the agent feedback and logs of agents using the QM CLI.

The first iTunes music app for Mac and Windows had an amazing feature. A user could type the name of a song, album, or artist, and in milliseconds their library would be filtered to show all of the songs that matched.

I loved this feature. Today I've added to QM a substring search across all files that I think is unique. It focuses on clearly communicating the location and filenames of content that matches, rather than the passage in the document. I think it's clearer.