r0b0tlab/hermes-concurrent-agents

Hermes Concurrent Agents is a Shell-based toolkit for running multiple AI agent workers simultaneously on a single machine with unified memory hardware, such as certain NVIDIA GPUs or Apple Silicon chips. The goal is to dramatically increase the total speed of AI task processing by taking advantage of how these chips handle memory and computation. The problem it solves: normally, running one AI agent at a time leaves GPU processing capacity underused. On unified-memory hardware, the underlying inference engine (the software that runs the AI model) can batch multiple requests together more efficiently, meaning that running three or four agents at once can deliver two to three times the total output speed compared to running just one. Each worker runs in an isolated environment and plays a specialized role, creative writing, coding, research, quality assurance, or task orchestration. Workers pick up tasks from a shared kanban board (a simple to-do list stored in a database), which handles task dependencies, prevents duplicate work, logs everything, and automatically recovers if a worker crashes. You would use this if you have a powerful machine with a unified-memory GPU, want to run AI agents on it locally, and need to process many tasks in parallel. It is aimed at technically experienced users who want to maximize throughput on local hardware rather than relying on cloud AI services. The project is written in Shell scripts and requires a compatible AI inference backend and the Hermes Agent tool to be installed.