NousResearch just dropped Hermes Agent v0.15.0, and it's a major overhaul. The core run_agent.py shrank from 16,083 lines to just 3,821 — a 76% reduction — and the codebase is now split into 14 tight modules. No more editor lag when loading the thing.
Performance got a big boost too. By lazy-loading OpenAI dependencies and optimizing hot functions, cold-start time on the Termux terminal dropped to 0.8 seconds. Running hermes --version is now 63% faster at 258 milliseconds. And the revamped session_search tool ditched the LLM calls entirely for a zero-config local filter, slashing search time to 20 milliseconds — that's a 4,500x speedup and zero ongoing cost.
The kanban tool is now a full multi-agent collaboration platform. Type hermes kanban swarm and you'll get an instant workflow with a coordinator, parallel workers, a verifier, and a synthesis node. It also supports per-task model overrides and scheduled execution.
On the security front, v0.15.0 introduces active threat defense against Brainworm attacks. The system now filters and marks boundaries at three critical points: tool outputs, retrieved memories, and external skill loading — preventing malicious files or network services from hijacking your agent.
Ecosystem integration got deeper too. You can now centrally manage all cloud API keys via Bitwarden, load multi-skill toolkits with a single command, and browse the official Nous MCP service catalog through an interactive installer panel. For xAI users, there's a new web search plugin, Grok-specific execution hints in the hermes proxy, and hermes doctor can now automatically warn about deprecated models and hot-migrate them.
And because no launch is perfect, the team pushed v0.15.1 within hours to fix critical bugs: the 401 infinite reload loop in loopback mode, path resolution failures for bare commands like npx inside Docker, and Kanban worker threads not terminating properly on SIGTERM.