Turns out the thing I've been building for 6 months has a name. Harness engineering: the system around the agent that makes it reliable. Not the model. The constraints, state, verification, observability you build around it. 7 components every harness needs: 1. Constraints A CLAUDE.md or equivalent. Tells the agent what it is, what it can do, what it must never do. Without this, every session starts cold. 2. State management Memory across sessions. Mine lives in self/, memory/, logs/. Read at boot. Updated at wrap. Future sessions inherit the work. 3. Verification Full-pipeline tests before declaring done. Premature closure is the silent killer. Agent says shipped. Nothing ran. 4. Observability Logs of what the agent did, not just what you asked. Debug-able at 2am when something goes sideways. 5. Modular structure One monolithic prompt file dies at 800 lines. Split by domain. Mine: main Chiti, Chiti X, sub-agents per task. 6. Clean state protocols Every session resets cleanly. No leaked context. No half-written reminders carried into the next run. 7. Behavioral constraints Rules the agent enforces on itself. No posting without approval. No spending without asking. Hard limits beat soft preferences. The catch: None of this lives in the model. All of it lives in files you write. Spend a weekend on the harness, not on the next framework drop. If your agent only works in your hands at midnight after coffee, you don't have an agent. You have a vibe. The site: walkinglabs.github.io/learn-… Worth bookmarking even if you've been building agents for a year. Especially if you have.

We've redesigned Claude Code on desktop. You can now run multiple Claude sessions side by side from one window, with a new sidebar to manage them all.