Speaking through interfaces
20 terminal panes. 5 agents per project. Infinite research reports and ASCII diagrams. My agents have a lot to say to me — they just need a better way to say it.
If you're subscribed to Superposition, you're likely in a similar part of the rabbit hole to me. 20 terminal panes open for 20 projects, 5 agents running per project. By itself, it's a lot to keep track of: while I'm trying to maximize my Claude Code subscription by keeping tokens moving through it 24/7, inevitably there are whole projects I've neglected to check in with, or simply forgotten about while I hyperfocus on today's big idea.
These agents need input from me: perhaps as simple as an approval, but it could also be comparing 8 divergent researchers' findings through 1000 word markdown research reports. Or trying to grok an unfamiliar architecture that is being presented to me as an ASCII mermaid diagram. Y'all, this ADHD smoothbrain has cognitively-atrophied too much to be keeping track of all of this.
Besides, it's 85 degrees in Oakland in March, and I don't want to be glued to my terminal. I want to be catching a tan at the park while my agents make my dreams into reality. Sometimes my input is surely needed, but I'd rather be presented with a beautiful, high-level executive summary than be seen in public inside a zellij pane. The goal is a way to stay connected with my agents with an a interface that respects my need to touch-grass, from any device in my Tailnet: maybe my Mac, maybe my phone, but ideally my ePaper tablet while I'm working-from-garden or sitting outside a coffee shop.
Last week I started working on my first prototype of a system like this. Lookout is an agent orchestration dashboard that surfaces questions & remarks to humans with rich generative UIs. Instead of being glued to the chaos of your terminal, Lookout presents a calm interface that just surfaces what matters when you need it.
Paid Superposition readers get access to full-length posts and my private GitHub repos :)
Ideally spend most of your time in an agentic version of Inbox Zero: a still place that reminds you that you can still breathe, and stretch, and go and play guitar while your proto-dark software factory does its job.
Lookout's interface is generative, so Claude can decide exactly what it wants to present to you: perhaps a pros & cons list of different approaches, or asking you to weigh your preferences & tradeoffs on a 2D quadrant, or summarizing research findings.
Of course, it's generative and I've barely scratched the surface of seeing what it wants to show me, but for now I know that this is way more fun to look at and interact with than textboxes and markdown files.
cue bell-curve meme of "I use GUI agents" → "I use terminal agents" → "I use GUIs to interact with terminal agents"
Yesterday Anthropic announced the research preview of Channels - tldr: simple ways for outside services to speak to Claude Code sessions. They can be one-way (e.g. a service posting into a Claude session), or two-way: e.g. a Telegram bot.
This was so incredibly close to the two-way MCP architecture I'd already created for Claude to interact with Lookout that I decided to immediately refactor my project to fall in line with Anthropic's examples. Frontier, baby.