For the past year or so, I’ve been very interested in AI Agents. These AI-powered things are fascinating to me, both in how they can augment human performance, but also how they’re built and function.

I’ve likely mentioned in past writing, but I’m not the most organized person. And I like to context-switch often, working on half-a-dozen little tasks across a few different teams and products. Every week is a struggle to remember what I did the week before, and what I need to do in the week ahead. I see Agents as a solution to helping me with the mess. I don’t want it to necessarily do the work, but augment my ability to do the work.

I’ve tried all the big ones, including Grok Bot, Meta’s Muse, OpenAI’s Dot, Claude’s Duet, OpenClaw, and a few more that I can’t even remember. New ones seemingly pop up every day; just go search the app store and you’ll find a bunch.

I’ve also tried to build my own at least four times, and I’ve really tried to make OpenClaw work. When Fable 5.2 was recently announced, I figured it was time to try again. I had gotten annoyed at OpenClaw’s bloat and the struggle to keep it alive, so I figured it was a good time to revive a past attempt.

Meet Bond, my latest experiment. This is really the 4th iteration of Bond. This latest one is built on Pi, and runs as a process on my machine with a selection of “clients” for interacting with it, including a CLI, Web UI, and macOS app. In reality, the foundation is incredibly simple. Bond lives in a folder on my machine, and stores all of its data as human-readable files. (It’s important to me that I own my data, and it’s one of the reasons I keep trying to build a personal agent.)

Bond mk4 has been alive for nearly two weeks. While planning with Fable 5.2, I started with a set of simple principles for how Bond should work and act.

  1. We build small, we use it, and then fix what hurts.
  2. We build CLI and GUI equally.
  3. We use plain, human readable files for data.
  4. We build and test safety and security systems.

The actual list was larger, but I think these are the ones that matter most. The first principle led us to build the Friction system. As Bond and I work, Bond will sometimes fail, or do something unexpected. Thanks to Opus 5.5 (Bond’s default model) the agent is actually pretty good at noticing most of its own mistakes as they happen and will automatically file a Friction note. But I also can tell Bond to file a Friction if I notice something I don’t like.

Bond will periodically review these Friction notes and offer to smooth them out. Bond was built with a plugin architecture, and a core system. The core system is fairly small and doesn’t do a ton on its own outside of chatting. Everything else in Bond is a plugin. And Bond has honed a process (after lots of trial and error) for self-improvement on the fly.

Bond’s basic system allows for one main chat that runs for eternity (in theory). But after a few hours of a single chat, I quickly realized I needed ways to spin up other channels, just to help organize and context-switch. So I mentioned this to Bond, and a few minutes later it restarted itself and I was able to start creating new chats with Bond.

It definitely wasn’t a very smooth experience to start; the first time Bond restarted itself, it did not come back online. But as we continue to work together, Bond’s systems have matured and its restart flow is much better now.

While watching the Ravens somehow lose to the worst team in the NFL (the Chargers) I was chatting with Bond on my phone. We worked towards improving the way channel icons are displayed, and ended up building this fun crystal orb. While Bond is working the orb’s glass effect distorts; it’s fun to watch while the agent works. And I’m looking to add different shapes, to let Bond morph as it sees fit.

Bond’s orb, while it’s working.
sphere
cushion
pill
puck
gem
drop
A few shapes I’m trying.