The story this week is agents growing up. Not just doing tasks, but remembering context, learning from prior sessions, and gaining better access to live data. Research and open-source tooling are converging on the same question: what does it take for an agent to operate reliably across a full workflow, not just a single prompt? Today's digest covers what's moving on GitHub and what Anthropic's research team published over the last few days.
rea: Reverse Engineering Anything with Agents
morluto/rea is a TypeScript framework for agent-driven reverse engineering — covering everything from app behavior analysis down to native binaries. It supports a broad range of targets including Android APKs, C/C++ binaries, JavaScript, Python, Rust, and Java, and integrates with Claude Code via MCP. The idea is that reverse engineering is fundamentally an investigative process: you form hypotheses, test them, and iterate — which maps well to agentic loops. With over 442,000 stars and active pushes today, it's among the highest-starred developer tools on GitHub.
hermes-agent: The Agent That Learns from You
Nous Research's hermes-agent is a Python-based AI agent framed around continuous improvement — it refines its behavior based on accumulated context and feedback rather than starting fresh each session. It supports Claude, GPT, and other models, and has become a reference implementation for teams building personalized agents. Pushed earlier today and sitting at 261,000 stars, hermes-agent reflects the broader shift from one-shot assistants toward agents with persistent, evolving state.
github.com/NousResearch/hermes-agent
firecrawl: Web Data Infrastructure for AI Agents
firecrawl/firecrawl has evolved from a scraping tool into something closer to a web intelligence layer for AI agents — handling crawling, extraction, and markdown conversion so agents can consume live web content without bespoke parsing code. Its maintainers describe it as "building the library for superintelligence," and at nearly 200,000 stars it's become a standard dependency in many agent pipelines. Active pushes today suggest the team is still shipping fast.
github.com/firecrawl/firecrawl
Anthropic: The Components of Agent Memory
Anthropic published a research post on November 6 explaining how Claude decides what to remember across conversations — and how a background "dreaming" process reorganizes stored memory and corrects inconsistencies over time. It's one of the clearer public explanations of episodic vs. semantic memory in production AI systems, and directly relevant to anyone building agents that need to operate over long time horizons.
Anthropic: How 207,000 Research Sessions Reveal AI's Role in Science
Also from November 5, Anthropic released an analysis of 207,000 research sessions to understand how scientists actually use Claude in their work — not how they say they use it. The findings inform how AI could meaningfully accelerate scientific discovery, with particular attention to the difference between automation (doing tasks faster) and augmentation (finding directions researchers wouldn't have considered). It's one of the more data-grounded takes on AI and research productivity published recently.