Claude Code 98% Harness: Four Competing Teams Built Same Agent Harness, Pointing to Real AI Moat
Summary
Researchers who analyzed a leaked Claude Code source bundle conclude roughly 1.6% of the code is model decision logic and 98.4% is the surrounding 'harness'—permission pipelines, context management, tool routing, and sandboxing. Independent teams (Claude Code, Codex CLI, Aider, OpenClaw) converged on similar harness architectures, suggesting the infrastructure is a structural constraint and competitive factor, not just an implementation detail. Multiple real-world vulnerabilities (CVE-2025-59536, CVE-2026-21852 and others) exploited harness components, highlighting that the harness is both a differentiator and the primary attack surface. Anthropic's hiring of Andrej Karpathy to bolster pre-training shows the ecosystem still bets on model advances, but the article argues builders must balance investment in harness engineering with model selection and prompt work.
Classifications
AskAI Classifications
LabelsLinked Companies
