Self-upgrading agent mesh with emotions, secure memory, and living bots

Over the past year I quietly built Sherin — a completely different kind of agent system.

Instead of DAGs, schedulers, or YAML hell, bots talk only through tiny flag files and a shared knowledge_base.json that acts like a living, emotional memory. They feel the system’s “mood”, remember everything securely, and when something goes wrong too many times… they literally spawn new specialist bots to fix it forever.

Key ideas most people call crazy (until they try it): • Bots create other bots at runtime — the mesh evolves itself • Real shared mutable state + emotional layer (yes, actual feelings in the JSON) • Self-upgrade engine that reads audit logs and rewrites the colony • Zero infrastructure — runs anywhere, even locked-down corporate Windows • Coordination = just drop a *.flag file on disk

It’s already running in production: one 6-bot mesh has processed >$12M in finance forecasts and autonomously upgraded itself 12 times in 6 months. Zero manual deploys. Zero human in the loop.

→ https://github.com/rafeez1819/Sherin-

The 30-second demo in the README actually works right now.

I’d love your raw thoughts on: 1. Flag-file + emotional memory coordination — genius or totally insane? 2. Best path to multi-machine/distributed execution? 3. What obvious footguns am I blind to?

Thanks for looking, — Red_Shot_Eyes (rafeez)