Wavestone teardown dissects 11 coding agents across 4M lines of code

Wavestone AI Lab published an 83-page source-code study of 11 production coding agents, including Claude Code, Codex CLI, Gemini CLI, OpenHands and Aider — roughly 4 million lines of Python, TypeScript and Rust with a 90-day longitudinal diff. None of the agents uses a general-purpose agentic framework or embeds code retrieval into its context loop; 9 of 11 adopted skills/plugins and 8 of 11 adopted MCP.

DEV Community

Open source

Wavestone teardown dissects 11 coding agents across 4M lines of code

11 agents studied: Claude Code, Codex CLI, Gemini CLI, OpenHands, Aider and others
About 4 million lines of code plus a 90-day longitudinal diff
9 of 11 use skills/plugins, 8 of 11 use MCP
AUTOCOMPACT_BUFFER_TOKENS and POST_COMPACT_TOKEN_BUDGET constants are gone from binaries

#Claude#OpenAI#Google#MCP

AI agents refactored 300K lines of code in three weeks for $4,000

Experiment: GPT-6 Astra wrote 75,000 lines of useless code in 35 hours

Claude Code's 25,000-token MCP limit bypassed by 49,964-token result

Security

Wavestone teardown dissects 11 coding agents across 4M lines of code

AI agents refactored 300K lines of code in three weeks for $4,000

Experiment: GPT-6 Astra wrote 75,000 lines of useless code in 35 hours

Claude Code's 25,000-token MCP limit bypassed by 49,964-token result

GitSpawn: one line in .git/config turns coding agents into command runners