Original Summary

AgenticView is a Claude Code plugin that turns your coding agents into a little 3D office. You talk to Atlas (the manager), Atlas splits the work and walks it over to the right worker, and the workers edit your real project files: on whatever provider each agent uses (Claude Code sessions on a Max plan, Codex, GitHub Copilot, Antigravity, Gemini, or any OpenAI/Anthropic-compatible endpoint you add). Things in the video: first-person walk mode with real collisions, live desk monitors showing each agent's work, whiteboards per room, a workflow timeline of who handed what to whom, automatic failover when a provider hits its rate limit (the agent walks to the manager's desk and complains), designated work desks, a lounge where idle agents play rock-paper-scissors, and a layout editor for the floor plan. It's a normal Claude Code plugin (open source, MIT): /plugin marketplace add VoidCU/agenticView then /agenticview in any project. And here's the part I like most: the office built itself. The agents you see in the video wrote and reviewed their own code, and even this overview video was produced by the office's Producer agent in the Production Room. They maintain their own codebase. And here's the part I like most: the office built itself. The agents you see in the video wrote and reviewed their own code, and even this overview video was produced by the office's Producer agent in the Production Room. They maintain their own codebase. How it's built (for anyone wanting to do something similar): It's a plain Claude Code plugin: a local Node server (Hono + WebSocket) that Claude Code launches with a skill, plus a React 19 + react-three-fiber front end. The whole 3D office is procedural primitives, instanced into a handful of draw calls: no modelling tools, no assets. Each provider is just its CLI driven headless ( codex exec , copilot -p --output-format json , agy -p , gemini -p ), with a small adapter that maps its JSON/NDJSON stream to one common event shape (text, tool call, file changed, usage). If a CLI can print machine-readable events, it can be an "agent": that's the whole trick. The Max-plan integration never launches Claude itself (Anthropic's terms): instead, your own Claude Code session runs a /agenticview-work skill that long-polls the office over MCP and executes each task as a background subagent. The office writes .claude/agents/*.md files so every office agent is a real Claude Code subagent with its own model and restricted tools. Rate limits are handled by classifying each CLI's error text (quota vs crash vs auth) and failing over down a provider order; the agent visibly walks to the manager's desk to "report" it, which started as a joke and turned out to be the best debugging UI in the app. Why: I had agents on four different subscriptions and no single place to see who was doing what. Making it a game meant I actually watch it. Repo: https://github.com/VoidCU/agen

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/28 06:38:26