I’m just going to dump my whole agentic setup out here, because I see too many people missing giant chunks of this and i

You can read the original post here: https://x.com/jamonholmgren/status/2076001786700394610?s=52

There’s a strong practical streak in this setup, and honestly, that’s what makes it worth paying attention to. The post is basically a field report from someone who’s been living with agentic coding long enough to see where the rough edges are. Not the shiny demo version. The real version, where the agent needs guardrails, memory, review loops, and a way to recover when things go sideways.

The core idea starts with an AGENTS.md router. That sounds small, but it’s doing a lot of work. Instead of hoping the agent guesses the right docs, skills, or tools, the router points it somewhere useful. From there, a customized workflow doc like AGENT_WORKFLOW.md keeps the session grounded in the way the author actually wants work done.

Then comes the part many people skip, self-healing documentation. Every system has docs the agent is told to keep current, with summaries near the top so they’re easy to find later. That’s not cosmetic. It’s what keeps a project from turning into a pile of half-remembered instructions and stale assumptions. I’ve seen that pile. It gets ugly fast.

The rest of the setup reads like someone who has been burned enough times to stop trusting vibes. Agents run the app themselves. Tests are written and kept current. Linters don’t just complain, they fix things or hand the job off to another model. Reviews happen at multiple points, with different personas looking at the work from different angles, maintainability, security, performance, and domain context.

There’s also a nice operational detail here, the trace worksheet. If an agent stops midway, another one can pick up the thread. That’s the kind of thing that makes autonomous work feel less like a gamble and more like an operating system.

One related piece that fits well with this thinking is Tau, which also treats agent work as something that needs structure, not just prompting.

The bigger lesson is straightforward. Agentic coding gets better when you build for continuity, recovery, and verification, not just output. That’s where the work starts feeling durable. And a little less like improvisation.