🔥 Hot Repo: 63K Stars — The Agent That Teaches Itself New Skills

NousResearch's Hermes Agent is the 1 trending repository on GitHub today April 12, 2026 with 63,456 stars. The v0.8.0 release, shipped on April 8, 2026, merged 209 pull requests and resolved 82 issues — roughly the scope of a major version bump, packaged as a minor increment. What It Does Hermes Agent is an open-source, model-agnostic AI agent written in Python. It runs as a standalone persistent process rather than an IDE extension: you send it messages from Telegram, Discord, Slack, WhatsApp, Signal, or the terminal while it executes tasks inside a local shell, SSH session, Docker container, Daytona workspace, or a serverless Modal environment that hibernates when idle and wakes on demand. The defining capability is a closed learning loop. After completing a complex task, the agent autonomously generates a reusable skill — a structured Markdown file — and stores it in SQLite. Skills are retrieved in future sessions, and the agent is periodically nudged to improve skills it has used before. Introduced in v0.7.0 on April 3, the feature is described as "autonomous skill creation after complex tasks" with skills that "self-improve during use." The practical effect is that the agent's performance on your specific workflows increases the longer you use it. v0.8.0 Highlights 209 PRs Background process auto-notifications. Long-running tasks — model training runs, test suites, CI pipelines — can now fire a callback when they finish. The agent picks up results when they land, without polling, and continues working on other tasks in parallel. Self-optimized GPT/Codex tool-use guidance. The agent ran automated behavioral benchmarks against itself, identified five failure modes in GPT and Codex tool calling, then patched its own system-prompt guidance. The release notes describe this as the agent having "self-diagnosed and self-patched" its tool-use reliability — an example of the meta-learning architecture applied to its own internals. Live model switching. The /mode