A living book about agentic workflows, agent orchestration, and agentic scaffolding
by
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.
Label-filtered activation — Every issues.labeled workflow now declares the specific label(s) it cares about via the gh-aw names: frontmatter key. Previously, all five labeled workflows fired on every label event, burning ~25 wasted workflow runs per issue through the full agent infrastructure just to have the LLM check a label and exit. Now each workflow short-circuits at the GitHub Actions level before spinning up any containers.
Consistent lifecycle labels — The fast-track path now adds the assigned label before closing, matching the slow-track lifecycle. Queries filtering by assigned will correctly find all completed issues regardless of which path they took.
Real web access for the research agent — The research-pass workflow previously declared web-search and web-fetch tools that the Copilot engine doesn’t support. It now has Playwright (headless browser, no secrets needed), the GitHub search toolset (code/repo/issue search via the existing token), and an optional Tavily MCP server (AI-optimized web search for when you add the API key).
The issue pipeline is central to how this book maintains itself. Every wasted workflow run is real compute, and a research agent that can’t actually search the web isn’t really researching. These fixes bring the pipeline closer to what it was always meant to be: a lean, self-correcting system where each agent fires only when needed and has the tools to do its job well.
A few lower-severity items remain on the radar: pinning CI action references to commit SHAs, locking the Pandoc container tag, and reviewing the PDF auto-merge policy. These are incremental hardening steps that don’t affect the core workflow logic.
This update was itself identified, researched, and implemented through the kind of agentic review loop the book describes — fitting, for a book that practices what it teaches.
tags: