punkin-pi

@deepfates so here are the details: 1) we modified pi so compensate for models at the time being inadequate (this was 8 months ago) + experiments around instruction handling. the harness modification is open source: hdresearch/punkin-pi 2) i used to work there not anymore but it was a microVM company (i believe the offering still stands afaik and let me know of you wanna get in touch with the ceo) so we built agent images that boostrapped this harness as an agent on a VM with instructions on running it. you could resize vms etc as needed. 3) we then ran experiments around vm fleet coordination. these where long running fleets. they would run for weeks on end. we could throw tasks at it. tasks would run for hours on end. it was goated because we could also branch them so if we ever felt like we need to fork a task or experiment, we would snapshot the agentvm with its current state and branch the entire thing, continuing the harness into another direction. context wasn’t a problem because the vm state carries everything so the harness would know whats up. internally employees had vm fleets (like 80+ VMs for example) active for days sometimes weeks on end. and ofc ssh vm action lifecycle was layered on top of this harness as a pi-extension (separate from this repo). another harness that I ran an overnight experiment with last weekend was the prime agent after their rewrite release. its open source. pretty good at running groups of these subagents. on my end i hadn’t set it up to run in VMs per se just in terminal. but yea it did a thorough audit of our codebase. lastly if it was good enough for prime intellect to have prime agent rewrite itself in rust then I would say its pretty capable too. hope that helps.