Wavestone dissects 11 coding agents, finds the same 7 building blocks inside
Arindam_1729 · x · 2026-10-25
Wavestone AI Lab analyzed the source code of 11 coding agents, including Claude Code and Codex, and found capable agents share the same 7 building blocks — agent = model + harness:
- Loop: a basic while loop for fully replayable runs
- LLM layer: single prompt template up to provider-specific configs
- Tools: basic bash to dozens of typed on-demand tools
- Memory: raw history to persistent cross-session notes
- Safety: step limits to rules, reviewer models, sandboxing
- Orchestration: one agent vs. spawning parallel sub-agents
- Extensions: Skills, MCP, and more
Also quoted in the thread: the author open-sourced OpenDuet (MIT) — the proactive Claude Duet agent on open models. It briefs you each morning, watches your apps while away, and pings only when needed; anything that sends/posts/pays waits for approval, each bot gets its own open model plus an editable MEMORY.md. Built with Mastra, Nebius, and Composio.
More from coding & agent
- LangChain details the five components of building a 'company harness' — hwchase17 · 2026-10-25
- Harrison Chase: Multiplayer agents are way harder than single-player — hwchase17 · 2026-10-25
- SwarmOS turns GitHub Issues into contracts that coding agents claim and ship as draft PRs — sn2006gy · 2026-10-25
- Adaptive Skill Granularity Framework Routes Coding Agent Skills at Runtime, Hits 100% Success — dair_ai · 2026-10-25
- Study: LLMs rewriting their own ARC-AGI-3 harnesss gain just 2.8 points, not significant — omarsar0 · 2026-10-25
- Dev adds thread tags to agent sidebar to track orchestrators across machines — kevinkern · 2026-10-25