No Claws, Just Vibes

I’m not an OpenClaw kinda gal. I dislike Claude CoWork, Manus and Genspark; they aren’t for me, I get a little lost and the agents turn me off by the shear amount of noise they generate. Too extra, which describes a lot of agents these days.

Now, I use agents a lot. As I type this (I’m typing every word, promise), I have two code factories rolling on my linux box (this machine) and a vanilla Claude on my mac. Outside of writing code, however, I find most agents more annoying than anything else - a lot of time they either leave me with a sense that I’m floating without accomplishing anything (every slide deck and doc editor agent I’ve ever seen), are too thin (tightly bound to a given product) or the opposite, too extra (see above).

Productivity wise, I have typically do the following:

  • prodigious calendar usage
  • track project detail on a whiteboard
  • take notes with pen and paper (I go through a new notepad every 2 weeks)
  • checklists in a notes app (mostly personal stuff)
  • paste notes to myself in Slack when I need to switch machines
  • markdown docs for long running projects or research or useful notes (e.g. I have a doc with every Postgres incident, fuckup or tricky problem I’ve encountered and how I fixed it, ungodly useful, could sell a book)

In this paradigm, many things are intermixed; I might have a to-do item that I make a calendar event for so I don’t forget, sometimes my whiteboard becomes my daily checklist, occassionally my notebook becomes a long form thought experiment that I’m too lazy to type up but keep the old notebook around so that I can reference it later, my calendar can get overwhelming so the month view is stressful to look at and I end up making a list in the margin of my whiteboard. So it’s a jumble but held together in my brain (most of the time).

Months ago I started to think about what a personal agent for me would look like since I very much disliked basically everything I have used so far. This was the list that I penned in my notebook:

  • support comingled calendar events with checklists & one off to-do’s
  • high level view of how busy i am
  • do something with my hand written notes and whiteboard sessions (not gonna stop analog)
  • connect my long format notes with all the other (usually ephemeral) cruft
  • cross machine support but basically free and not trapped in someone’s app
  • fire and forget voice notes -> to-do’s
  • keep clanker noise to a minimum

So here’s what I came up with

I rebuilt this guy multiple times and ended up with what I think is a pretty good attempt at something my brain can make sense of. It is generally trying to not be noisy; any calendar event or to-do can have a spec, that’s where you hide the noise. There are artifacts; created by the agent in response to some to-do, etc - for example, I have been making invoices as artifacts from a to-do, then I combine that with “lists”. Lists are just lists of objects; you have to define the schema of it but they are linkable to to-dos, available to the agent, etc.

Also in service of not being noisy, the app UX was designed to not be full screen. That’s why it boots so smol. The chat window isn’t a sidebar, its a small panel at the bottom, because I want to agent to do things but I don’t want a long back and forth - I don’t really want to spend more time in the app at all, so it’s de-optimized for long agent conversations.

Books. Books are just collections of markdown files. I like books better than “notes” because I like to add structure and hierarchy to a collection of notes. Sometimes at least. Sometimes, like my daily standups, I just record these things to improve my recall ability. And I do that with pen and paper. Sometimes if I feel like I should keep whatever i jot down, I take a picture of it and hand it to the agent or I will read aloud to the agent and let it record it in a page in the blackhole of one of the books.

Then there’s the multi-machine requirement. The agent runs in a sandbox VM, why not use that? Sure, so the VM can sync from my linux machine where I do most of my work, to my mac or windows box. The VMs run on the machine itself, if I get around to putting this on a phone, I’ll put it in the cloud but I like the intentional syncing between devices for whatever that’s worth.

heyvm sync push todo-agent

What’s on the roadmap?

I have a rough sense that creating background agents that get managed via to-do’s might be something I want. I don’t have a use case that I’m dying to have but there are a handful of things that might be nice.

Pluggable agents. This is technically possible, as the app just inserts an agent into the VM and starts it up. But it’s a hack and not a first class experience by any means. Basically, separate the agent harness from the app.

Local LLMs. I want a nice agent harness around a LLM that runs on my hardware and costs nothing.

Batch processing. Things like “break this document up into 100 parts, verify for accuracy, and return sources”.