What developers are learning about agents
The revolutions in AI silicon, in models, and in large-scale cloud hosting are fascinating. But ultimately, they are only a small part of the story.
When we look back at the electrification of the country, it was driven by revolutions in large-scale generation and transmission, though not that many people were employed in utility-scale electrical infrastructure. But everyone found their life turned upside down by the pervasive applications of electricity. It affected businesses, residential life, urbanization, social lives, everything.
The AI revolution is similar. Most of us will never have anything to do with utility-scale cloud computing, models, or GPU silicon. But every single aspect of our work and home lives will be impacted by AI.
The work that software developers are doing to figure out how to wrangle agents is fascinating. While still pretty low in the stack, the knowledge gained about managing agents will affect all our work, our education, our hobbies, our healthcare, etc.
Everyone is sharing their agent experience, there is a ton of learning availalbe out there. Some observations this week that I found interesting:
Hot take: AI coding tools don't replace developers. They make systems thinking portable across any language or framework. – Dave Beckett (staff engineer, various places): The Harness is the Product
This jives with my usage. AI coding tools are not a replacement for developers — they allow developers to do more.
Hot take: The most valuable AI skill is no longer prompting. It's building the loop around the model. – Dave Beckett (staff engineer, various places): The Harness is the Product
It is not just a loop; it is the orchestration of multiple agents using the model. How do you structure work, how do you share state and guidance across agents, etc. The way you structure your work can have dramatic impacts on completion time and token efficiency.
This next article is a bit academic in spots but worth the read – humans still have a lot to add in choosing the problems to address, and in feeding the right context to the agents:
I believe this pattern generalizes. The two channels from the Brynjolfsson-Hitzig framework appear in practice as two distinct gaps where human judgment creates the most value. The first is problem formulation: defining the strategic question and specifying what data should exist to answer it. This is the Channel 1 gap: deciding what to codify and how to codify it well requires understanding the business deeply enough to know which questions are worth asking. The second is curation: assembling the right codified information for a specific AI task. This is the Channel 2 gap. It requires domain judgment about what the model needs to see and what it should ignore. In an economy where highly capable AI is a commodity, these two gaps are where knowledge work lives. – Chris Walker: Context Engineering: Why Hayek's Knowledge Problem Survives AI.
Steve Yegge always has pithy observations (From IDEs to Agents with Steve Yegge):
Meanwhile, “slot machine programming” – building 20 implementations and picking the winner – is becoming normal practice for teams. – Steve Yegge
I’ve been wondering if we would see this behavior and sure enough, people are starting this practice. When prototyping and building are cheap, it is only rational to build something several times and pick the winner.
The IDE could be evolving into a conversation and monitoring interface, not a code editor. – Steve Yegge
For years, the IDE was dumb as a stone, letting me make stupid mistake after stupid mistake. I love the new world of an agentic, conversant IDE.
All of these observations are primarily concerning software development and the IDE app. There is no productivity app moat, though. I certainly don't need to manually create spreadsheets anymore; I am happy to let an agent do the work, and I don't really care what platform the agent uses – Excel, Google Sheets, some custom Python code, whatever.
Shorts
I continue to poke around on my state tax comparison app using Claude Code. I am starting to conclude that the Washington state tax burden is not that bad. But the state isn’t spending the money well — for instance, Matt Yglesias points out that Washington state does not compare favorably to Tennessee in education results, and Tennessee has a much lower tax burden. I plan to start looking at comparative state spending levels, and then onto measures of performance by state.
Also from Matt Yglesias — Everyone gets canceled sooner or later:
You may as well just stir the pot.
An aging guy on loss of significance:
One of the things that they don’t tell you about retirement is that you stop being who you were and you basically disappear.
I feel sorry for this guy. None of us were ever all that important, and you can’t measure yourself by what others think of you. You have to create your own meaning and satisfaction.
Ben Thompson on Copilot Cowork, Anthropic’s Integration, Microsoft’s New Bundle ($). Ben is generally positive on Microsoft’s AI work. I find it to be a branding and usage train wreck. Copilot, Copilot Cowork, AI-at-Work, Work IQ, Microsoft 365 Copilot, E7, E5, Agent 365 — as a customer, I have no idea what all this means; as a developer, I have no idea what to target.
Microsoft may agree; they have just reorganized all the Copilot efforts into one team.
Torsten Slok: Margin Debt at Record Highs:
With margin debt at record highs, any downturn in stocks risks turning into a sharper correction as leveraged investors are forced to sell into falling markets.
We saw this in 2000/2001 as the dotcom era unwound.
Construction Physics on The Elusive Cost Savings of the Prefabricated Home:
However, these hopes have yet to bear out, and achieving cost savings with prefabricated construction has proved to be highly elusive in practice.
We tried to go down the prefab path; some of the available products look very compelling. But the schedule and cost benefits vanished. Permitting was just as drawn out and expensive, site work was just as extensive and expensive, and we still needed to hire a GC to handle everything. One would hope that AI would drive a dramatic improvement in permitting costs and permitting timelines.