harishkotra/agent-office

What it solves

AgentOffice creates a virtual, pixel-art office environment where AI agents can autonomously collaborate, manage tasks, and grow their team. It provides a visual and persistent way to observe and interact with multi-agent systems, removing the dependency on cloud-based LLMs by supporting local inference via Ollama.

How it works

The system uses a "Perceive $\rightarrow$ Think $\rightarrow$ Act" loop. Agents run a thinking cycle every 15 seconds, using an LLM (via Ollama or OpenAI-compatible APIs) to determine their thoughts, actions, and tool calls. State is synchronized in real-time across browsers using Colyseus, while a Phaser.js and React UI renders the agents as pixel-art sprites moving through a virtual office.

To maintain continuity, the project uses SQLite for persistent memory and Ollama embeddings with cosine similarity for semantic memory retrieval, allowing agents to remember interactions across sessions.

Who it’s for

Developers and researchers interested in multi-agent collaboration, emergent behavior, and local LLM orchestration. It also serves as a customizable engine for creating simulations like AI classrooms, virtual hospitals, or business simulations.

Highlights

• Autonomous Team Growth: Agents can independently decide to hire new team members (interns, devs, designers).
• ** لقب Local-First**: Runs 100% locally with Ollama, ensuring zero lock-in.
• Integrated Tooling: Includes sandboxed JavaScript execution, web search, and note-taking capabilities.
• Persistent Semantic Memory: Combines SQLite and embeddings for importance-weighted recall.
• Interactive UI: Features a drag-and-drop layout editor, a real-time system activity log, and a "Click-to-Follow" focus mode for agents.