Multi-Agent Loop (VS Code)

A VS Code extension that runs a file-based Planner/Writer/Checker loop using GitHub Copilot (VS Code Language Model API). Each agent keeps its own memory file and hands off to the next agent via a DECISION block.

Quick start

1. Run Multi-Agent Loop: Start
2. Fill .multi-agent/task.md or use Multi-Agent Loop: Chat
3. Planner → Writer → Checker runs in order, each agent writes to its own file

File layout

The .multi-agent/ folder is created automatically:

• task.md: task/goal
• planner.md: planner memory + history
• writer.md: writer memory + history
• checker.md: checker memory + history
• state.json: loop state
• chat.json: requirements from the chat panel

Settings

• multiAgentLoop.modelVendor: default copilot
• multiAgentLoop.modelFamily: optional
• multiAgentLoop.modelId: optional
• multiAgentLoop.autoContinue: true/false
• multiAgentLoop.maxCycles: max steps per start
• multiAgentLoop.memoryMaxChars: prompt memory limit
• multiAgentLoop.historyEntries: number of history entries to include
• multiAgentLoop.agentFolder: agent folder name
• multiAgentLoop.outputLanguage: output language (auto follows task/chat)
• multiAgentLoop.chatTheme: chat theme (auto/light/dark/sunset)

Notes

• If Copilot access is unavailable, model selection will fail.
• Agents must end their output with a DECISION block; the extension uses it to pick the next agent.
• When task.md, chat.json, or agent files change, the file watcher triggers the next run.
• The Multi-Agent Loop: Chat panel writes requirements into chat.json.