A living book about agentic workflows, agent orchestration, and agentic scaffolding
Latest updates and news about the Agentic Workflows Book.
February 2026 may be the month when agentic coding stopped being a niche and became the default way serious software gets built. Here is what the landscape looks like right now, why it matters, and where the pressure points are.
Today we hit a class of failures that are easy to misdiagnose: not prompt bugs, not repository logic bugs, and not deterministic workflow bugs. They happen in the integration gap between GitHub APIs, hosted runner networking, and third-party artifact delivery.
We just shipped a round of fixes to the agentic workflow pipeline that processes community suggestions filed as GitHub issues. These changes make the system significantly cheaper, more correct, and better equipped to research topics.
The updated Agents for Mathematics and Physics chapter now covers the extraordinary 2025–2026 breakthrough in mathematical agents—AxiomProver, Aristotle, Numina-Lean-Agent, AlphaEvolve, and more. But a question raised during the writing process proved more interesting than any benchmark table: why do unsupervised AI conversations never venture into mathematics or physics?
In the chapter note, Claude asks a sharper question than “models need tools”: if models can reason in language, why do they not naturally drift toward mathematics and physics discovery in open-ended conversation?
We just expanded Skills and Tools Management with a new section on the Model Context Protocol (MCP) and how it shapes modern agent tooling.
We just added a new section on GitHub Agentic Workflows (GH-AW), the emerging approach that turns markdown instructions into secure, composable GitHub Actions powered by AI agents.
Welcome to the Agentic Workflows Book - a living, self-maintaining guide to building with AI agents!