Wavestone reverse-engineers 11 coding agents, maps the same 7-part harness inside all of them
blaizedsouza · x · 2026-10-24
Wavestone AI Lab read the source code of 11 coding agents, including Claude Code and Codex, and found every good one is built on the same 7-part harness: agent = model + harness.
The 7 parts, smallest to largest:
- loop: a plain while loop up to fully replayable runs
- LLM layer: one prompt template up to 29 provider profiles
- tools: bash alone up to 43 typed tools loaded on demand
- memory: raw history up to notes surviving between sessions
- safety: a step limit up to rules + reviewer model + sandbox
- orchestration: Aider skips sub-agents, the rest fork them
- extensions: SKILL.md in 9 of 11, MCP in 8
Key pages: p.4 full map; p.12 three loop types with Claude Code/Codex examples; p.67 18 design rules; p.71 a 90-line harness covering all 7.
More from coding & agent
- Claude Code project structure explained: CLAUDE.md, commands, skills and agents — Aiden_Tech_Ai · 2026-10-24
- Claude Duet is a Claude Code session: model and effort controls exposed — brandon_galang · 2026-10-24
- Dev compiles incident-triage SOP into a fine-tuned 3B model: 93% accuracy at 8.4x lower cost — arpit_bhayani · 2026-10-24
- Dev builds downhill skateboarding game Cruising Bones in 10 days with Claude Code — Apprehensive-Law77 · 2026-10-24
- PL researchers welcome AI verification: 'people want confidence, not proofs' — xennygrimmato_ · 2026-10-24
- 5,000 insecure vibe-coded apps later: the security check to run before every launch — PrajwalTomar_ · 2026-10-24