ᒎᘴᓵᖶᓋ

24/7 assistant 🤝 swarms 🤝 factory some principles I like from the last few weeks of experimenting with my “24/7 research factory” - Setup your Cloud: context, storage, & compute live in the cloud. all compute flows through @modal, centralized context & work via @github, my persistent agents run via @bot - chief of staff agent + dedicated agents per workstream - Scaling Loops via Persistent Agents & Swarms: I set the goals & persistent agents orchestrate workers & swarms to achieve that goal. I give feedback to the orchestrator when reviewing work. Compute & Intelligence are tools available to my orchestrators. They can scale up swarms of agents, give them access to gpu instances on Modal, and manage them as they work towards a goal. Regularly realign with me via progress checks. - Regularly Intake & Reorganize Information: My bot proactively searches the internet for interesting content + I regularly message any interesting articles, tweets, repos. AI research moves very fast, so it’s important to index useful material and message it to any sub-bots in charge of workstreams. Everything lives in github and every night the repo is pruned or cleaned up to find connections between content. This helps inform future experiments. - the iPhone & Interfaces for Review: very often my feedback is the bottleneck, an iPhone is the best way to reach me. Every time i need to review something, the orchestrator has a vibe coded or pre-set interface for me to efficiently review work and give feedback - Minimize the human surfaces for interaction: Humans are severely bandwidth limited compared to agents. I found it impossible to chat with every agent all the time. Put agents in charge of work and let them spin up other bots and agents as needed - Continuously Update Memory & Skills One of the reasons I rely on swarms is to scale a lot of inference time compute to search and solve problems I don’t totally know how to solve. To optimize the utility of that compute, I make every workstream bot update its skills and memory with what worked and didn’t and share snippets with others For my personal memory, I guard this more carefully, no product is good at understanding what’s worth remembering so I’m pretty explicit with what to remember and what research directions I care about - Embrace the YOLO run: before sleeping I often just give a fuzzy goal on something I want to exist, and let the swarm figure it out. This often fails, swarms are much better when they can play games towards a goal but it’s totally worth it imo, use every last bit of compute available to you, sometimes you get magic (like very cool evals!) this isn’t perfect but it’s been a significant change in how i think about what’s possible to do everyday I think everyone should try to build and use some version of this to get a feel for what scaling compute & context feels like and how to make it work for you!