8 Comments
User's avatar
Daniel (Dig. Craft Workshop)'s avatar

This piece made me stop and hold my own setup up against yours, and I'm grateful for that. I run a single assistant with a shared memory and session rather than a fleet, so a lot of your coordination machinery solves a problem I don't have — but reading how you think about it was clarifying in its own right.

Two things genuinely opened my eyes. First, your separate adversarial reviewer that's only allowed to find faults and never to fix them — I don't have that clean split between the mind that writes and the mind that criticizes, and I can see why I should. Second, the "what would look broken while still reporting fine?" lens on monitoring, with that SSH tunnel running 21 hours to a dead address. That one's going to change how I look at my own watchdog.

Thanks for showing the actual wiring instead of just the outcome. Rare and useful.

Ahad Amdani's avatar

That's great to hear that it's helpful, Daniel! Just trying to apply the lessons I've learned around observability and the practical problems I've encountered running multiple agents (especially around the same piece of work).

If it helps save anyone some time or trouble, sharing my setup is worth it, especially given that the harnesses themselves are evolving and not necessarily in line with each other.

I plan on sharing the "meta-harness" I've come up with as well, so that you can fall back from your primary LLM driver (and model, even) while still getting consistent results, regardless.

I know you have the Reels product that you are building out. That's been built with a single agent? Impressive!

Phil Powis ❤️⚡️'s avatar

Love the family dinner part. Can't get those back.

Karo (Product with Attitude)'s avatar

That was a really good read. Thank you so much for sharing, I always enjoy learning how others set up their agents. Do you have any agents that use Discord to talk not only to you but to each other too?

Ahad Amdani's avatar

I'm actually figuring out how I want to approach that next. I use Gbrain as my knowledge store, and have it "dreaming" but haven't cracked all of its use, including the minions concept and the ability to manage agents and message across them. Additionally, OpenClaw as a possible route to having the agents talk to each other. I have been wanting to upgrade the fleet, but honestly using Fizzy and having them self-manage has been working out well enough for all of my workflows. I even recently got the Hermes setup with Qwen3.6 working as a good enough fallback, and have even gotten a MacBook air setup with gemma4 quant to do local inference and software development. So as far as having them talk to each other on Discord, I want it but haven't needed it due to the agents having discussions the Fizzy cards. I would like to have it all there at some point though because it's a faster message than my dispatches, which can take a few minutes to respond.

Karo (Product with Attitude)'s avatar

Very interesting. When you get there and write about it, please let me know. That's something I haven't tried yet either.

Patrick Lefler's avatar

This is really good - quite helpful.

Ahad Amdani's avatar

Appreciated, Pat! I'm putting together the companion post with all the technical bits spelled out and additional details, but hopefully this content still gives you all of the how, even if not the exact implementation. Hope the results of building your own are as outsized as mine, and more!