Wavestone dissects Claude Code, Codex and 9 more agents, finds a 7-part harness blueprint
blaizedsouza · x · 2026-10-05
A Wavestone AI Lab study tears down 11 coding agents including Claude Code and Codex and concludes that an agent is a model plus a harness, with every harness built from the same 7 parts that differ only in scale.
- Loop: Mini-SWE-Agent is a plain while loop; OpenHands logs every event for replayable runs
- LLM layer: one prompt template plus 29 provider profiles
- Tools: from bash-only to 43 typed tools loaded on demand
- Memory: from full-history retention to agent-maintained cross-session notes
- Safety: from step limits to policy rules + reviewer model + OS sandbox
- Orchestration: Aider deliberately skips sub-agents; others fork them with separate contexts
- Extensions: hooks, skills and MCP
The companion article, How to Design an Agent Harness, distills six decisions that turn a model into a worker you can leave alone, targeting pain points like agents claiming done without running tests or forgetting rules mid-run.
More from coding & agent
- Ponytail Hits 150k+ Stars Telling AI Coding Agents to Stop Overbuilding — we93 · 2026-10-05
- Dev predicts companies will wall off MCP servers, ushering in agent-to-agent era — JosephJacks_ · 2026-10-05
- Dots and Muse AI Coding Tools Can't Resize Browser Below ~500px, Blocking Mobile Testing — pkragthorpe · 2026-10-05
- Dev orchestrates 6 AI skills for a test-fix pipeline, spending just $3/week on DeepSeek — Own-Awareness8037 · 2026-10-05
- fastapi-gql-mcp: one GraphQL MCP surface cuts tool tokens from ~11k to ~2.5k and stops payload bloat — tangkikodo · 2026-10-05
- Tencent Hunyuan's RSR Boosts 27B Model Terminal-Bench 2 pass@3 from 57% to 74% — Tencent-Hunyuan · 2026-10-05