🔥 Hot Repo: 172K Stars — The Agent That Outran Codex CLI

One-liner — Nous Research's self-improving AI agent just shipped v0.15 with its 16K-line core shrunk 76%, session search made 4,500× faster, and a cold-start benchmark that now beats OpenAI's Codex CLI head-to-head. - Repo: NousResearch/hermes-agenthttps://github.com/NousResearch/hermes-agent - Stars: ⭐ 172,390 +est. 1,200 today - Language: Python - License: MIT --- What It Does Hermes Agent is an open-source, self-improving AI agent from Nous Research that runs on any infrastructure — $5 VPS, Docker, SSH, or serverless via Modal or Daytona. The core hook is a closed learning loop: the agent creates skills from experience, improves them during use, and builds a persistent cross-session model of who you are. Switch between 200+ models via hermes model with no code changes; talk from Telegram while it runs on a remote VM. Why It's Blowing Up v0.15.0 "The Velocity Release" shipped May 28 with numbers that don't usually appear in a single release: 1,302 commits, 747 merged PRs, and 321 community contributors — the project's largest release ever. The headline refactor compressed runagent.py from 16,083 lines to 3,821 -76% across 14 cohesive modules, with behavior identical and every test patch path preserved. The practical payoff shows in benchmarks: hermes --version now flips the head-to-head cold-start comparison against OpenAI's Codex CLI, with 47% fewer per-conversation function calls shaved off along the way. The release also lands on the same week Anthropic shipped Opus 4.8 — and hermes-agent's GitHub topics include claude-code and anthropic. v0.15.0 adds Opus 4.8 support and ships a Nous-approved MCP catalog with an interactive picker, giving it a direct upgrade path for developers already running Claude Code. The Kanban multi-agent platform — built across 104 PRs — now supports full swarm topology: hermes kanban swarm creates a root, parallel workers, gated verifier, gated synthesizer, and shared blackboard in a single command. A Brainworm-class promptw