Team-Wide Agentic Harness

Ian Langworth describes his emerging role as a "Harness Guy" — someone who optimizes the environment around AI coding agents so the team gets more consistent, correct results. He runs brown-bag lunches where colleagues trade tips and he demonstrates skills (such as one that manages a Linear queue, generating customer-tailored release notes).

Skills are files, so they're inherently shareable and check-in-able. What's harder to share are the conventions those skills depend on: sandboxing rules that scope all agent work to a single directory, keeping dotfiles, email, and package installation out of reach, with standard directories for worktrees, plans, and temporary artifacts.

Langworth argues the whole thing — conventions, skills, context documents — should live in a version-controlled "harness" directory. He draws on The AI-Native Startup Handbook, which recommends a shared repository of context, plugins, and skills across an organization. An evergreen docs directory describes the company, product, and recurring procedures — the atomic blocks of context fed into a context window before dispatching an agent.

Because skills are code, they deserve PR review: another set of eyes before they change everyone's agent behavior. Langworth closes by wondering whether conventions that work comfortably in one person's head will survive contact with a full team.