On the Web on joe blubaugh

My AI coding setup today

Thu, 16 Apr 2026 04:38:34 +0000

My AI coding setup has been evolving the last few months. Here are some things I’ve tried and moved on from:

Gas Town. It’s just too much to wrap my head around right now.
Pi and custom harnesses. I’m not at the stage where I need to deeply customize my workflow, starting from zero. I’m also not at the stage where I’m frequently switching models for different tasks. The Claude CLI is still good for me at this stage.

Here’s my current simple-to-use setup today:

Claude Code CLI. I use a variety of plugins with Claude Code. I don’t use IDE integrations or the Claude desktop app.
- superpowers provides a basic software development framework that really improves the quality of initial implementation. I don’t even think about this one, it gets invoked automatically all the time.
- autoresearch provides a simple optimization loop for measurable tasks. I recently used it to cut down bundle sizes in one of the programs my team ships.
- Slack’s official plugin for reading messages.
- gdocs-cli. Anything at work that doesn’t happen in Slack happens in Google Drive.
- Grafana-internal skills and tools for our wiki and queries.
- bugbot-loop listens for PR comments and responds to them by implementing good suggestions and resolving bad ones, over and over.
Cursor BugBot installed in our Github repositories. This is the tool that I think provides the right combination of ease and high quality in its review processes. Combining this with bugbot-loop immediately increases PR quality without intervention on the part of any humans.
agent-deck. A Tmux layer that provides management of multiple agent sessions. It understands git, especially worktrees, and watches the state of sessions. This way I can tell which of my sessions is still churning, and which are idly waiting for intput.
ars contexta. This one is a little complicated to explain, but it’s a knowledge base that an agent automatically maintains and refines over time.

So, my work setup tends to look like this: I have a custom skill for summarizing Slack that runs once a day via launchctl. It DMs me a summary each morning that helps me build up my to-do list. Throughout the day I throw my notes into ars contexta. I used to use Logseq for this, but I’m doing less manual creation of notes now.

I have an organized set of sessions in agent-deck.

A section of coding tasks - each associated with a set of PRs. When the PRs are finished to main they’re killed and cleaned up. After the initial coding, when the PR is updated I start /bugbot-loop for up to 10 iterations.
A set of longer-lived “research” sessions where I’m keeping notes about and getting feedback on larger ideas and direction for projects. These usually end up writing markdown files that are consumed by ars contexta or written into things like Architecture Decision Records for a given project.
a single ars contexta session where I throw in notes. I can also ask this to search recent Slack activity or gdocs for information to add to notes.

As far as concepts and tools I’m interested in investigating to continue improving my workflow:

agent-deck “conductors” to increase autonomy for task-oriented sessions even more.
Archon. Providing a little more determinism in my software workflows could be nice. Compare “fix a bug” to “explore the codebase for improvements”. These practices should probably look different, and . Seems like a promising way to increase autonomy with some of these tasks.
Spec-driven development. I need to dedicate more time to reading and prototyping here. I’m optimistic about improving autonomy in my sessions via more detailed planning.
Context graphs. The concept feels like it may be able to encode a lot of “Chesterton’s Fence” information that agents struggle with. In particular I’m interested in it from an Observability standpoint. Every running software system is unique, and we’ve never been able to establish industry-wide practices that everyone follows to the letter. Semantic Conventions and the like are moving Observability in that direction, but discovering structure is somewhere that LLMs are very promising for.

Zen: A Browser You Can Love

Sat, 07 Feb 2026 01:17:18 +0000

Everyone is putting AI in their browsers now. Good (maybe even great!) browsers have been discontinued so users can focus on AI.

I’m not opposed to using AI agents for all sorts of things, but I don’t feel comfortable giving it nigh-unfettered access to the web. I’m not ready to let LLMs take on life-altering responsibilities like booking flights, looking at my banking sites, or reading my emails. I don’t trust their data retention policies with potentially sensitive context like calendars or emails. Nobody’s

So I don’t want it in my web browser, which is where most of this work gets done in 2026. I’ve been using Firefox at home and Arc at work, and they’ve got different problems. Arc is over. “The Browser Company” has gone all in on an AI Browser that’s not dissimilar to OpenAI’s browser. Firefox added AI features in November and is only now adding controls that can give you some control over data sharing.

Fortunately, someone recommended Zen to me. Arc’s approach to tabs always clicked with me: every view of a profile, even in a different window, has the same tab set. You can use multiple profiles (“spaces”) to manage different tab groups, and different sign-in contexts. Zen’s latest release uses the same behavior, and it’s delightful. It has a simple but very useful split-view feature. I use it for things like reading the meeting notes and having the video call in the same window, or comparing two related documents without having to tab back and forth. Where it has any AI features at all, they are limited and require advanced settings to turn on. You’re not going to get them intruding on you, at least for now.

It turns out I’m not the only one who’s had this experience.

Multi-agent coding and the resurgence of the terminal

Thu, 05 Feb 2026 15:23:52 +0000

All of the exciting AI coding is happening in terminals. Codex is good, the various VSCode forks are good, but if you’re doing multi-agent orchestration you’re generally using terminal UIs, roped together with Gastown, agent-deck, OpenClaw, lemonaid, or something else.

It’s gratifying, as someone who never left the terminal for a “grown up” IDE, to see how small programs piped together is the paradigm for innovation here. The stuff that makes it possible? Simple, largely open-source tech:

tmux: lemonaid, agent-deck, and Gastown all use tmux panes and windows to let you jump to the right place and give the right info to the right agent at the right time.
git: Lots of work-tracking and communication systems (like beads) use git. Git worktrees are an essential component for multiple agents to work on the same codebase at the same time.
sqlite: agent-deck uses this for communication, and so many other tools seem to as well.

Introducing nvim-beads: Manage beads in neovim

Wed, 14 Jan 2026 06:29:35 +0000

A few days ago my co-worker Bob turned me on to Beads. Beads is a step up from things like PLAN.md when managing work for coding agents, because it provides more structure, is easy for agents to parse, and provides things like progress tracking for multi-agent workflows.

I was intrigued by the setup because it seemed like a way to improve my agent’s ability to work autonomously, but I needed a way to try it out without stepping on top of some complex multi-person projects at work. I decided to use bd and a coding agent to implement a neovim plugin for managing beads.

The result is nvim-beads. Implementing this took about 106 commits over 3 hours of pairing between me and the Claude CLI. It’s written in technologies I know about but am not good at: Lua and the Neovim API. There’s no way I would have built something with this level of complexity in this environment in less than a week of good, solid work. It has tests - which I wouldn’t know how to write.

I don’t think it would have gotten built without Beads, either. The workflow of building a PLAN.md and then having an agent execute it works OK, but if the plan is large, you’re wasting a lot of context using it over and over again. Beads’ structure wastes less context on large plans, and make it easy to either blow away an agent and start again, or makes conversation compaction less of an issue. It also makes parallel agent workflows on multiple git worktrees easier to manage, because the agents can share beads state on one beads-specific branch while doing their work implementing those beads on feature branches.

I think the tech is worth trying, and this post from the author Steve Yegge is a good place to start. After that, read the best practices post.

Podcasting is dominated by warm applesauce

Tue, 13 Jan 2026 11:37:44 +0000

The end of podcasting as a viable cottage industry really does seem to be here now, and the end of podcasting as a culture-moving media format has arrived. Amy Poehler won the first podcasting Golden Globe for a “hang out and promote my new project” podcast. Most celebrity podcasts are this way. Even the ones that ostensibly have a theme: Smartless is barely anything anymore. They’re warm applesauce, and they dominate the format to the point that ad revenue is draining away from a lot of better shows.

“Actors give podcast award to actor with a podcast” isn’t all that shocking, but still, look at the nominees for “best podcast”:

Good Hang with Amy Poehler
Armchair Expert With Dax Shepard
Call Her Daddy
The Mel Robbins Podcast
SmartLess
NPR’s Up First

I guess we can at least be grateful they didn’t nominate JRE. For every single niche these shows occupy - maybe with the exception of Up First - you can easily find a better version that is less popular. I guess the fact that “massive media breeds mediocre media” is maybe not all that surprising, either.