Harness Anatomy: a cross-cutting view of eight agent harnesses

Eight production-scale agent harnesses, looked at side by side. Not as frameworks to learn, but as specimens. The point of the page is the heatmap — what each harness chose to build deep, what it chose to leave thin, and what the cross-cuts say about the field as a whole.

This page sits beside the Dual-Axis Framework, the Pattern Evolution map, and the Paradigm Evolution timeline. Together they ask one question at four scales — paradigms (how engineers think), patterns (named solutions inside a paradigm), the matrix (the structural anatomy of the agent paradigm), and now harnesses (specific implementations of that anatomy in 2026). The matrix is the map; the harnesses are the territory.

How to read this page: the heatmap below is the centerpiece. Rows are harnesses, columns are cognitive functions, cells are coverage strength on a 0–3 scale. Click any framework name or any cell to open a side panel with the framework's main-loop location, positioning, and where it goes deep. The list is not closed — the same method extends to any new harness.

Coverage heatmap — 8 harnesses × 7 cognitive functions

Read a row to see one harness's profile. Read a column to see which capabilities the field has matured (Action is universal) and which are still being built out (Governance is the 2026 investment frontier). Cell shade encodes engineering depth, not feature checkbox.

The five-step method that builds the heatmap

The heatmap above is not opinion — it is the output of a fixed procedure applied to each harness's source. Five steps. Roughly two to three days per framework, reliably reproducible. Every cell in the matrix is anchored to a file path and a line range; the depth shade comes from how that code is shaped, not from how the framework markets itself.

The five steps are the method. The walk-throughs — actually applying these five steps to Claude Code's tool registry, Aider's git loop, DeerFlow's planner-researcher-writer split, Codex CLI's seatbelt sandbox, and four more frameworks, with file paths, line numbers, and the reasoning behind every classification call — are the substance of the Harness Anatomy training course. The web page tells you what to do; the course is the on-site reverse engineering done together.

Five shared engineering ancestors

Look across the eight harnesses and the same five engineering ideas keep surfacing. None of them are new. None of them are agent-specific. They are what software engineering accumulated across fifty years for problems that only look new from inside the agent bubble.

Three takeaways the heatmap makes visible

First: every harness has a main loop, and they are structurally the same — a probabilistic version of a reconciliation loop. Second: the field has standardized on Action; the variance is everywhere else. Third: the harnesses with the widest coverage are not the dev-tool ones — they are the personal-AI and multi-agent ones, where the requirements force every cognitive function to be present.