🚗 Main Branch: The One Where Agents Learned to Drive

Hiya friends,

Last week the lobster got loose. This week the agents learned to drive.

GitHub shipped Agentic Workflows and it is exactly the kind of thing that sounds like AI marketing until you see it 😉. Maintainers are already doing a lot with it: analyzing thousands of open issues and surfacing what matters.

---

🚢 What Shipped

GitHub Agentic Workflows (Technical Preview)

Define repository automation in Markdown. A .github/workflows/workflow.md file is all it takes. The gh aw CLI compiles it into standard GitHub Actions, executed by Copilot CLI, Claude, or Codex. Workflows run with read-only permissions by default. Write actions (creating PRs, issues) go through safe outputs that are reviewable and controlled.

We needed a way to let agents do real work in production repos. Agentic Workflows is that bridge making AI capable of maintaining a repository. It can triage issues, investigate CI failures, update docs, and propose fixes, all within a visible, auditable, and sandboxed environment. Peli de Halleux built an Agent Factory showcasing 150+ specialized agentic workflows. Worth browsing.

Copilot Coding Agent: Zed, Windows, and Model Picker Expansion

Copilot in Zed hit GA, so you can use your existing Copilot subscription. Copilot coding agent now runs in Windows environments via custom runs-on in copilot-setup-steps.yml (use self-hosted runners or Azure private networking since the integrated firewall isn’t compatible yet). And Business/Enterprise users can now pick between Claude Opus 4.6, GPT-5.3 Codex, or Auto per task instead of being locked to Auto.

Breaking free from GitHub Discussions’ limitations - by Jamie Tanna

He’s a maintainer on Renovate and wanted to know which feature requests his community cared about most. That’s not a big ask. But the Discussions UI has no way to answer it. Same for “which threads haven’t gotten a reply in weeks” or “how many bugs did we close this month.” Basic maintainer questions. So he built his own pipeline. Go to hit the GraphQL API. SQLite to store 11,000 Discussions and 55,000 comments. Evidence for visualization when he needed charts. The whole thing syncs every morning and gives him a triage dashboard that loads in 0.01 seconds.

---

🧰 What I’m Using

I posted a tweet asking about the most underrated tools with generous free tiers. 200+ replies later, here is what kept coming up:

Tool (Link)	What It’s For
Tailscale	Private Networking (mesh VPN)
Headscale	Self-Hosted Private Networking
Cloudflare	Secure Tunnels & Zero-Trust Access
Infisical	Secrets Management
Bitwarden	Password Management
Appwrite	Backend as a Service
Convex	Reactive Backend
Netlify	Frontend Deployment
Cloudinary	Media Management (images/videos)
Grafana	Observability & Monitoring
PostHog	Product Analytics
CodeRabbit	AI Code Reviews

The full thread is worth a scroll if you’re building anything on a budget. I wish all companies would learn: the free tier is a feature, not a compromise.

---

✨ This Week

Was streaming with @elbruno and spotted a slick tool he used for live demos. Meant to ask what it was, totally forgot. So I built my own: annotation-overlay. Shipped and open-sourced it by EOD. Made it for me, but if you do live coding, demos, or just want to draw on your screen, you might like it!

Building your own thing is delightful. AI makes the building part faster than ever. But I think it’s worth keeping an honest perspective: building is easy now, maintenance is still hard. Mitchell Hashimoto said recently that “there should be way more forks”, and he’s right. Sometimes a solution just for you is the right scope. Ship it, use it, move on.

---

That’s the week. Agents are getting real jobs. The free tier is a feature, not a compromise. Fundamentals first. Always.

With gratitude, I’ll see you next week,

Andrea

🇪🇸 Léelo en español

Support Main Branch: Subscribe | GitHub Sponsors