A worked reference, not a framework

An agent you can hand work to.

The roles, routines, hooks, memory, and task coordination that turn a coding agent into a system you can trust to make progress while you're away.

One person's actual setup, redacted and published with the reasoning attached. The example runs in Claude Code, but the architecture ports to any agent runtime. This page is the five-minute tour; the docs hold the depth.

The shape

Five layers, top to bottom

From how you talk to it down to the agents that do the work. Select a layer to see what it holds and why it's there.

The pieces

The modules

Each is a cohesive cluster with one upgrade boundary. Tap one.

The thinking

Nine load-bearing patterns

If you read one thing, read these — each as problem, pattern, why it beats the obvious alternative, and what it costs. Expand any card.

They're not independent

Three instincts, each applied in two places. Keep damage out of durable surfaces — the credential law and the file-protection hook. One source of truth, referenced not copied — the roles library and memory hygiene. Route by consequence, not by confidence — classify-then-act for tasks, tier-by-impact for findings. Adopt each when you feel the friction it removes, not before.

In motion

A task, end to end

What static files can't show: how one task moves through the system without you babysitting it. Step through it.

Go deeper

Read it, fork it, or hand it to your agent

The docs follow Diátaxis: why, what, how, and scaffolding to copy.

Hand the repo to your own agent Tour this repo. Read PATTERNS.md, then META_ARCHITECTURE.md, then WORKFLOW.md, then scan samples/. Summarise the patterns most applicable to my workspace.

Standing this up inside an organisation?

I'm James Ross. I design agent workspaces and AI-orchestration systems, and this is the reference version of my own. If you want these patterns adapted to your stack, the practice site is jamesross.ai.