The Harness Is the Thing: Why Your Agent Rig Matters More Than the Model

A solo developer argues that the decisive tool in agentic coding is no longer the model but the harness — the shared scaffolding of skills, an AGENTS.md file, and TUIs that sits between a developer’s intentions and an LLM’s raw capability. By keeping that layer unified across Cursor, Claude, and a tool called Pi, the author has effectively commodified the underlying models: switching providers becomes a non-event, and cheaper models handle routine work while frontier models are reserved for planning-heavy features and large refactors.

The workflow formalizes agent roles rather than trusting one prompt to do everything. Exploration produces an explicit DAG of tasks; a worker implements it node by node; a critic simplifies and challenges the result, often kicking work back to the worker; and a final ‘promoter’ step forces the author to actually communicate what was shipped. Splitting planner, worker, and critic into isolated roles — borrowed from Can Bölük’s ‘prewalk’ handoff technique and an advanced-harness playbook — is credited with cutting frontier-model usage in even intense sessions by about 75%, keeping the whole operation affordable on two $20 subscriptions.

The piece also shows the harness reaching outward: the author’s simple camera-to-vector app is scripted via a headless-Chrome driver so agents can generate posters and animations without burning tokens on repeated runs, and File System Access API integration lets the harness sync notes, config, and creative projects. The broader claim is that a well-built harness gives a single developer team-scale output and genuine agency over how solutions get engineered — while still leaving room for humans to keep the fun, creative parts for themselves.