In Person vs. Over the Phone — A SuperClaw Slack Bridge Retrospective
The baseline and the goal
The idea was simple: leverage a well-established social platform — Slack — to enable chatting with my SuperClaw agent without having to be on the box. The bridge went live on August 31, and from that moment SuperClaw was one DM away, wherever I happened to be.
SuperClaw from anywhere
This enabled me to interact with SuperClaw from anywhere I can use Slack — like on my phone while I am away from the lab. The box stopped being a place I had to sit at and became something I could simply talk to.
Building it while using it
It should be emphasized that we built this integration while using it at the same time. It continues to be a learning experience: finding out when I could leverage Slack to get tasks done, and when I needed to not use the bridge so I could be sure it wasn't part of the problem at hand. Debugging a transport you are actively talking through has a hall-of-mirrors quality — sometimes the honest move is to step out of the room and go look at the box directly.
Although the interaction was always virtual, using the bridge is like working remotely, versus being on the box, which felt like being "on location" or "in the office/lab".
Stabilization, and the "in person" distinction
Next came the stabilization of the bridge — making it more resilient to crashing and restarts at any level, in any part of the integration chain. The first days delivered two identical socket-mode crashes and the watchdog that now resurrects the bridge when it dies quietly. Then there was the adding of new features, such as the title argument to the /new command — which itself created additional issues that needed to be resolved.
Through all of it there was a definite behavioral distinction: realizing that I couldn't use Slack for some things. It was like needing to be there "in person" versus "talking over the phone". Some work only happens on location.
The division of labor
I tried to have the agent do all of the "grunt" work as much as possible — only needing to own and execute those manual steps that the agent couldn't or shouldn't own. Examples: management/account-wide API keys; actual commands and observations on systems or environments not accessible from the sandbox; and the need — the wanting — for me to "do" some of the things myself.
The collaboration was sometimes one of guidance (in either direction), and other times just outright command and control.
What I gained
The conclusion should be one of wonder: I have gained a colleague, a collaborator, a subordinate. But have I created additional work for myself? There is an overhead cost to doing it this way. For now I am convinced that I am:
- saving time — finding out how to do things while executing the steps;
- feeling more interactive — despite it being only with an AI/machine, not a real person — and feeling like I am learning about myself in a way that can only be done via collaborative efforts;
- bringing dissimilar projects together and accomplishing more — the whole becomes greater than the sum of its parts; and
- having fun!
Because of what I gained, I wonder how far this could go — and I am excited to find out.