The projects gaining traction on GitHub today reflect a shift toward agents that are persistent, cross-platform, and continuously improving rather than stateless one-shot tools. Developers are no longer asking whether they can build an agent — they're asking how to make agents that survive session boundaries, operate across every environment their users inhabit, and get better over time without manual intervention.
NousResearch/hermes-agent: An Agent That Learns From Experience
Hermes is a self-improving agent framework from Nous Research that builds skills from its own experience and recalls past conversations across sessions. It connects to 200+ model providers, deploys across six infrastructure backends (local, Docker, SSH, Modal, and more), and surfaces in Telegram, Discord, Slack, WhatsApp, Signal, and the terminal — meeting users wherever they already work. A built-in cron scheduler and subagent spawning make it capable of running autonomous background workstreams, making it one of the more complete production-grade open agent frameworks available.
github.com/NousResearch/hermes-agent
affaan-m/ECC: A Performance System for AI Coding Harnesses
ECC is a large-scale optimization layer for developers using AI coding assistants — covering Claude Code, Cursor, Codex, and others. It ships 63 specialized sub-agents, 251 reusable skills, 34+ language-specific rule sets (TypeScript, Python, Go, Rust, and more), and 15+ hook automations that fire on session start, build events, and security scans. The goal is to eliminate the friction of adapting AI workflows to each new tool or project: instead of reinventing conventions from scratch, teams inherit a research-first, production-tested system that has evolved over 10+ months of real-world use. With over 208,000 total stars and 1,300+ added today, the demand for structured AI coding discipline is real.
CopilotKit/CopilotKit: Generative UI Across Every Platform
CopilotKit picked up another 360+ stars today, reflecting continued momentum for its approach to agent-native frontends. The framework lets developers build chat interfaces, generative UI components, and full agentic workflows across React, Angular, Vue, React Native, Slack, and Microsoft Teams from a single backend. Its useAgent hook gives UI components direct read/write access to agent state, and an early-access continuous learning feature automatically tunes agent prompts based on real user interactions — closing the feedback loop between deployed agents and the humans using them.
github.com/CopilotKit/CopilotKit
Panniantong/Agent-Reach: Eyes Across the Open Web
Agent-Reach gives AI agents the ability to search and retrieve content across major platforms — social networks, news aggregators, and video sites — without requiring paid API access to each. The premise is practical: agents that can only see what's exposed through official APIs are operating with blinders on, and integrating broad web retrieval at the agent infrastructure layer is more sustainable than patching it into each individual application. At 21,800+ stars and still climbing, it signals real demand for affordable, wide-coverage information access in agent pipelines.