Claude Code: 1.6% model. 98.4% harness.

AI Engineering · Architecture

36one.cloud / Fractional Architect Architecture note · 002 / May 15, 2026

v1.0 · AI Engineering · Architecture

Claude Code: 1.6% model. 98.4% harness.

What Anthropic actually shipped, according to the source-level analysis by MBZUAI and UCL — and why this changes where your team should invest.

AUTHOR Myriam Spitz Mooser Fractional Architect — Microsoft 365 & Azure AI
FOR Tech leads & VP Engineering

Reading ~ 5 min · 9sections

02— THE MISCONCEPTION

Architecture, not the model.

Models are the innovation. The whole stake: pick the best LLM. The rest is plumbing.
— The shortcut to avoid
  • Anthropic invested in deterministic infrastructure, not decision scaffolding. The harness is the product edge — and it reveals what the model alone cannot do.
  • 01 — Architecture, not the model
  • 02 — The supervision paradox
  • 03 — Human atrophy
03— THE NUMBERS

Precise analysis. Not intuition.

Researchers from MBZUAI and UCL published a source-level analysis of Claude Code on arXiv, titled Dive into Claude Code. They went through the codebase line by line.

1,884 files analyzed TYPESCRIPT
512,000 lines of code the full codebase
1.6% AI decision logic the model
98.4% infrastructure (harness) everything else
04— THE HARNESS

What the 98.4% does.

Six mechanisms the model alone would not invent. The harness does not complete the model — it makes it operational.

  • Permissions

    7 modes, ML classifier.

  • Compaction

    5-layer pipeline, context preserved.

  • Native tools

    54 tools, real execution.

  • Hooks

    27 event types, plugins & policy.

  • Persistence

    Append-only session, sub-agents, recovery.

  • Isolation

    The model never touches FS, shell, network.

SOURCE Liu, Zhao, Shang & Shen — Dive into Claude Code (arXiv, April 2026)
05— 01 — ARCHITECTURE

The bet: infrastructure, not scaffolding.

The paper states it explicitly on page 3. A foundational choice, not a technical convenience.

The architecture invests in deterministic infrastructure — context management, tool routing, recovery — rather than in decision scaffolding.
— Excerpt · page 3
  • The bet: increasingly capable models extract more value from a rich operational environment than from frameworks that constrain their choices.
06— 02 — GOVERNANCE

Consent is no longer a guardrail.

The harness compensates for the model's flaws. It does not compensate for the atrophy it induces in the operator.

The harness locks down dangerous actions; it does not create the vigilance that knows how to say no.
07— 03 — TEAMS

The harness does not offset what it degrades.

He et al. 2025 — Cursor adoption across 807 repositories. The velocity gain disappears after the third month. The complexity remains.

Velocity gains collapse at month three. The harness does not cover the operator's loss of judgment.
— Synthesis
08— TAKEAWAY

The edge is the operator.

09— SOURCES

Sources & references.

Original LinkedIn article and studies cited in the analysis.

— PRIMARY SOURCE

— CITED STUDIES