Let Claude run 100+ agent campaigns over months, pointed at one end goal.

Earlier this year I had Claude run a 10-week long agent orchestration campaign, with ~200 agents working sequentially to verify some of my math research in chaos theory, eventually producing publishable output. I realised the the agent orchestration workflow my campaign was running on was relatively sophisticated, and could function towards any goal, or as a persistent team, so I made it a repo anyone can use. I used my original folder with the entirety of the tested 10w campaign as the foundation, and generalised it such that anybody can work with it. The idea behind splitting it up into orchestrator, wave general, and subagents in different terminals is such that the Wave Generals can absorb the context window that the subagents will produce, and can only give the crucial details to the orchestrator, while recording everything in .md files. This allows the orchestrator terminals to run for a lot longer. The workflow records everything. The subagents, wave generals & every wave have premade plans, frozen exit criteria, and handbooks to allow them to follow protocol. The subagents produce a NOTES.md , the wave general produces a wave_report, and a the orchestration then autonomously launches the next wave, and so on. Everything is planned beforehand, and verified live. One of my favourite features is it's ability to survive rate limit interruptions, and now with Anthropic's recent update allowing terminals to talk to each-other, this repo becomes even more powerful. hope this can bring about some really great projects, allowing for a better way of working with claude agents! SuperTeams: github.com/RajveerKapoor/SuperTeams/blob/main/README.md