Talk prep ask: how do you use (or think about) AI agent harnesses? Also migrating an orchestrator from bash to Elixir

Chat & Discussions AI / LLMs
, ,
1

Hi all :waving_hand:

I’m presenting “Building a Harness for Fun and Profit” at Colorado Startup Week on September 14, 2026.

tl;dr - I built a harness that lets AI coding agents work together and am going to hold a workshop/demo for people who are interested in doing this on their own. It’s got Claude Code and Codex CLI pairing in adjacent tmux panes over a file-based message protocol, coordinated by a small Elixir daemon, with Gemini doing real-time OTel telemetry analysis in anotherr tmux window. The thesis of the talk is that the magic isn’t necessarily the model, it’s the harness: e.g. a file protocol, a small daemon, and tmux can beat clever prompting.

A couple of reasons I’m posting here:

  1. The Elixir angle. I’m open-sourcing the stack before the talk and I’m migrating the orchestration layer (currently ~1,300 lines of bash driving multi-round agent collaboration) to Elixir. This includes stuff like task lifecycles as gen_statem, on-disk envelopes as the durable state, all side-by-side with the bash reference implementation. It’s early and honestly very experimental, but if OTP-shaped orchestration of AI agents sounds fun to you I’d love some eyes on it once it’s public and will have contributor-friendly issues ready.
  2. The research ask (the main one). I want the talk to answer real questions and not just demo my setup. If you use agents in your workflow – or deliberately don’t – I’d love a short virtual call to chat about things like how do you coordinate agents today? What would make a harness worth adopting vs. building yourself (the build vs. buy debate)? What would you be skeptical of? What would you ask in the Q&A?

If you’re willing to chat, please send me a DM and I’ll send you a link to my Calendly. BTW I’m located in the US MT/Denver time zone.

I’ll happily share back a summary of what I learn and the repos when they land in this thread.

Thanks!

2 Likes
2

Sadly I have nothing to contribute: I use bare Claude Code and only gradually adding hooks to slap the LLMs back to their senses when they make all the same mistakes (i.e. try to use rg like they would use grep, or ignore a higher-level tool for finding function callers / callees and still try to use rg/grep instead, and multiple others).

Super interested for the results of your interviewing community members, though!

I’d love to read or hear a summary when you believe you’re done gathering info.

1 Like
3

@dimitarvp , thanks for your comment! I’ve got several calls lined up and plan to update this thread when I’ve gathered some useful info. I also plan to build this in the open while I migrate the orchestrator from bash to Elixir, so stay tuned!