Agents and other topics

Claude Code and ChatGPT excel at short-lived projects.   Whip up a prototype app.  Summarize an issue for me.   Look at these documents and answer this question.   Wham bam, they are done.   

As I use agents more for long-lived projects, things get a little harder.   I have to apply more effort and discipline in formulating plans and objectives and defining success, more effort in monitoring the process, and more effort to ensure that intermediate steps don’t go astray.   The UXs of Claude Code and ChatGPT aren’t great for managing a long-term multistep process.

And when I shut my laptop screen and put my device to sleep, progress comes to a gradual halt.   I really don’t want the controller for my long-term projects and all the agents to be resident on my device.  I want my laptop to be a window onto my workflow, which persists in the cloud.  

I could go off and figure out how to stand up various things in the cloud and manage them, and maybe I should do that.   But I’d be happy if someone did this for me.

Roo Followup

I had a chance to talk with the Roo Code founders this past week; Brad Silverberg introduced me.  Great guys, their story of how they came to where they are today is fascinating.  

One of the things they talk about is the Roo cloud — running agents in the cloud, accessing them from anywhere.  I am very interested to see how this rolls out; as discussed above, I want to get to the point where my laptop is just a window on my persistent tasks in the cloud.

I also loved how they are using Roo Code themselves — their Roo agents automatically engage on each GitHub issue and automatically propose fixes.   It is Roo Code all the way down!

And I love this.  When you build a tool for yourself that you use and depend on every day, the rate of improvement can be dramatic, because you just can’t stand to live with the warts.  I am always drawn to teams that build the tools they need for themselves, whether it's for developer scenarios or end-user scenarios.  Every day, these teams get a chance to hone their product and hone their judgement, and products and teams get better and better.

For the same reason, I’ve never been drawn to enterprise software teams, and there is a reason why enterprise software is often clunky and hard to use — it is nearly impossible for the development teams to self-host on it, and so the flywheel never takes off.

I’m excited to see where Roo goes with this.  Benedict Evans commented briefly this week:

Old: generative search
New: generative content 
Next: generative product

It does feel a little like an avalanche is starting — agents using agents to improve themselves, where does that end up? 

Good science

It is good that we are maybe starting an avalanche of productivity in software, because we are screwing up in the sciences more generally.

I came across this Isaac Asimov article this week on the notion of right and wrong when it comes to science.   It is a clear articulation of how science is never perfect, but it gets better and “righter” over time, which is the nature of the scientific method.  

Our current government is throwing aside science in favor of political orthodoxy and cronyism – defunding science investment, suppressing undesirable results, eliminating scientists from government roles, and elevating pseudosciences to peer levels.

Cristi sent me this great video by Jessica Knurick explaining how 20th-century Germany went down the path of politicizing science, elevating pseudoscience, and purposely turning its back on scientific best practices.  It did not turn out well, and the parallels to today are all too clear.

Geographic apportionment is dumb

Part of the problem with our government is due to the way apportionment works in the US, and bad behavior on this topic is on full display these days.

Why do we apportion representatives by geography (and weird, convoluted geographies)?  This made sense once upon a time, maybe.  But local issues are well handled by city, county, and state representatives — I don’t need federal representatives to deal directly with these issues.   For federal issues, my local affinities are not strong.   On federal issues, I have more in common with my fellow SIC code employees or with my age group than I do with residents of Federal Way or the International District.  

You could apportion by last name, or shoe size, or hair color, and it would be about as meaningful as geographic apportionment.  Apportionment by age group or income might make even more sense and is less susceptible to gerrymandering.

Or, better yet, move to a proportional representation scheme.  Doesn’t require any constitutional changes, is not susceptible to gerrymandering, and is arguably fair.  

And as a strategic approach, it makes much more sense for the Democrats to embrace a scheme that eliminates the politicized gerrymandering process and get out of the tit-for-tat gerrymandering war with the Republicans.