In my last update on Hydra I explored some limitations in what I’d built. Honestly I didn’t fully explore the potential adjustments. Something felt wrong about the entire project. It has languished.
I did make the change I promised there. I let the Control loop run multiple turns in the same session, and it helped. The agent got less confused. But plenty of struggles remained. Hydra kept losing track of what it was working on. Most importantly, it struggled with the concept of handing off work to other agents. Every agent wanted to do all the work. The sandboxing was a mess, which probably didn’t help. It leaked too much visibility, and the roles never seemed clear.
What I loved about Hydra was the concept of a small kernel that enables it to build itself. I ask for capabilities (such as communicating over Telegram) and it builds that. The architecture is simple, pluggable and customizable. Without even touching the core, Hydra could become whatever I wanted it to.
Hydra was supposed to be an OpenClaw clone—focused on the realm of an AI personal assistant. Communicating over varied channels, operating software for me. Send some emails, browse the web.
But day-to-day I was using it to build itself. This is just a software project. All the struggles were the common struggles of building software. Project management. Some bug tracking. But most important of all was deployment. Hydra at one point started hallucinating release procedures and deployment systems, submitting its fixes to deployment queues that didn’t even exist yet.
So now I’m taking a different approach. An approach that seems to be all the rage these days. We’re all building AI-enabled software engineering platforms. I took Gastown as nearly a cautionary tale of AI psychosis, yet I think there is something here. What has changed is what always changes: the models just get better.
I’m keeping the same idea I loved about Hydra. My new coding factory is a small core that is extendable. The difference is that the core is focused on being a software development platform rather than an agent cognition platform. It’s not just AI agents talking to each other to solve arbitrary problems. It’s AI Coding Agents building and releasing software.