Skip to content
ID
← All work
Daily toolAI2026

Multi-agent dev harness

My own development team of AI agents, with shared memory, test gates and voice input.

A harness on the Claude Agent SDK that runs a small team of role agents: producer, planner, designer, scribe, triager, retro and a code agent. Their shared memory is a plain Obsidian vault of markdown.

The same harness is cloned per project. It drives Unity C#, Python and TypeScript work, including this website.

My role

Solo developerDesigner and sole developer

Team

AI Lab · 2026

Platforms

  • Windows

Stack

  • Python
  • Claude Agent SDK
  • MCP
  • SQLite
  • Obsidian
  • faster-whisper

Solo developer

What I did

  1. 01Designed and built the whole harness on the Claude Agent SDK.
  2. 02Role agents with a shared Obsidian-vault memory and a reviewed memory-candidate queue.
  3. 03Build and test gates the code agent cannot bypass.
  4. 04A local code index, local-model tooling and on-device voice dictation.

What stands out

  1. 01Memory is reviewed rather than appended: agents propose memory candidates, and a scribe promotes only what stays true.
  2. 02Hard gates: the code agent cannot end its turn while the build it touched is red.
  3. 03A local code index is queried before broad file reading, and a local model answers questions about big files at zero API cost.
  4. 04Voice dictation into the prompt, fully on-device with faster-whisper.

See it in action

Demo video coming soon.