fifthrevision

  • Essays
  • Blog
  • Projects
  • Bookshelf
  • About

Blog

From my head to yours, no filter.

  • November 2026
  • October 2026
  • September 2026
  • August 2026
  • July 2026
  • June 2026
  • January 2026
  • November 2025
  • June 2025
  • April 2025
  • March 2025
  • February 2025
  • January 2025

Release: Axle 0.35 and 0.36

For some reason, Axle releases often come in pairs. Often one release ships something new and interesting, then it prompts follow-up work. In a way, that’s the mark of a productive design iteration.

0.35 highlights

  • Continuations. A send can now carry after, a program that runs once the turn settles. This holds the queue and allows follow-up messages. This is useful in places where a program wants to run a judge process in loop.
  • Cost tracking. Axle now has computeCost, and when paired with ModelCatalog, can compute cost priced from models.dev. This handles context-length pricing correctly as well, and can price across an entire Agent run.
  • Providers and Protocols. One long-awaited change is to split protocols from providers. Now you can run the Anthropic Messages protocol against Bedrock, or OpenAI Responses API to OpenRouter.
  • Ergonomics and performance. Finally, we cleaned up errors so that client-side handling is simpler and more coherent. Also long overdue, multiple tool calls now run concurrently.

0.36 highlights

  • Remote Agents. The client-server layer every agent app rebuilds is now in the library. A numbered event log on the server, a RemoteAgent in the client with the same API as a local Agent, and resume after a dropped connection built in. It’s the design we’ve been working toward for a long time, and it does most of the heavy lifting.
  • Deferred tools. Create tools that can be handled later. Deferred tools do not require an execution block when instantiated so they are perfect for scenarios such as asking the user for inputs. Runs remotely too.
  • Agent lifecycle events. To round it all out, agents now fire lifecycle events so that you can observe and respond to agent changes. This is useful for running parallel programs and saving agent state.
Posted Nov 11, 2026

Intellectual Labor

I think a lot about the nature of work, especially how it pertains to my own. What it comes down to, put simply, is whether one believes work to be zero-sum.

As best I can tell, building software is decidedly not zero-sum. Jevons paradox applies here, but I’ll go even one step further: there will be not just more consumption, but compounding productivity.

Software benefits from specificity; it really sings when tailored to a person, organization, or workflow. When building software is expensive, countless hours – our work – are spent on debating and front-loading decisions in hopes of reducing the expensive building, not to mention everyone else working around the limitations that are deemed too expensive to fix.

When building is cheap, however, work can go into judging whether the software performs to our satisfaction, and more importantly, into making it work well. I see this in my own work: instead of slogging through boilerplate or speculating on interaction patterns, I can build, judge, and iterate far faster than before. I don’t think it diminishes the value of work in this case, even if it diminishes the margins of software contracts.

I’ll posit that software productivity gains unlock at different levels. Right now, we gain productivity from software but lose some of it to sharp edges and workarounds. Soon, these edges will be filed down and gaps filled, which will then give us time to address higher-order concerns. That will give rise to new opportunities and perils as we build upon or reconfigure systems that were built on an earlier paradigm.

But that’s how expertise works, isn’t it? Once the foundations are laid, one is free to wander around the broader space to solve other problems that crop up. These lead to often unforeseeable breakthroughs that move the collective forward. I’d like to think that our intellectual labor works that way too, not by carving a fixed pie up between us and AI, but by leveling up and growing it as our competency grows.

Posted Nov 06, 2026
  • Blog
© 2009–2026AI PolicyI speak to computers, telling them about dreams that humans dream.