Outworked tackles the challenge of managing multiple Claude Code agents by providing a visual orchestration layer that breaks down user goals into specialized subtasks and routes them to configurable agents. This macOS desktop app offers a Phaser-powered office interface where agents operate as animated sprites, communicate through a shared message bus, and integrate with external tools and services. If you’ve ever needed to coordinate multiple AI agents with different roles and personalities, Outworked offers an observable and manageable approach.
what outworked does and how it is built
At its core, Outworked is a multi-agent system for Claude Code, designed as a macOS desktop application. It transforms Claude Code from a single-agent environment to a visual, multi-agent orchestration platform. The interface is built on Phaser, a popular HTML5 game framework, which powers the “office” where agents appear as sprites with roles, personalities, and models.
The architecture centers on an orchestrator component that decomposes a natural-language user goal into smaller subtasks. These subtasks are assigned to specialized agents that can run different Claude Code models and have distinct roles configured by the user. Agents communicate through a shared message bus using a straightforward syntax: [ASK:AgentName], which allows them to request input or data from one another.
Under the hood, each agent has full access to Claude Code’s tools like Bash, Edit, and Read, enabling complex workflows that combine code execution, editing, and data retrieval. Beyond Claude Code’s native capabilities, agents can interact with external services and systems via MCP servers, messaging channels like iMessage and Slack, a built-in Chromium browser, and Cloudflare tunneling for external network access.
Additional system features include scheduled triggers using cron syntax to automate tasks, a skills plugin system defined by SKILL.md files for extending agent capabilities, SQLite-backed persistence for state and data durability, and permission gates to control sensitive operations securely. The asset pack system allows customization of the visual appearance of the office and agents, including backgrounds, fonts, furniture, and role-specific sprite sheets.
the multi-agent orchestration pattern and technical tradeoffs
Outworked stands out by implementing a multi-agent orchestration pattern where a natural-language goal is decomposed into parallel subtasks routed to specialized agents. This decomposition and routing logic is the system’s technical heart. It enables complex workflows to break down into manageable pieces assigned to agents best suited for each subtask.
The inter-agent communication pattern using [ASK:AgentName] is a simple yet effective message bus approach that avoids complex shared state. Agents send and receive messages via this pattern, allowing asynchronous collaboration and task handoff. This design reduces coupling and makes adding or modifying agents straightforward.
The permission gate system is a thoughtful addition that secures calls to sensitive tools like Bash or external services. This prevents accidental or malicious execution of risky commands, which is essential when agents have broad capabilities.
Using Phaser for the visual office interface is both a strength and a tradeoff. It provides a clear, interactive visualization of the multi-agent system with animated sprites and a customizable environment. However, it ties the UI to a specific platform (macOS desktop), potentially limiting portability or web-based deployment.
The architecture cleverly combines AI agent workflows with traditional application patterns such as SQLite persistence and cron scheduling, balancing AI flexibility with robust engineering. The plugin-based skills system encourages modularity and extensibility, allowing users to introduce new capabilities without altering core code.
Code quality in the TypeScript codebase is reportedly solid, with clear separation of concerns between orchestrator logic, agent definitions, and UI components. The project leverages Claude Code tooling extensively, making the entire system effectively an extension and orchestration layer rather than a standalone AI engine.
quick start with asset packs
The simplest way to get started with Outworked is by creating an asset pack folder with just a PNG sprite sheet:
~/.outworked/assets/my-pack/
default.png # employee sprite sheet — that's all you need
For more customization, you can add a manifest and additional assets:
~/.outworked/assets/my-pack/
manifest.json # optional — everything auto-detects without it
background.png # replaces the office background
myfont.ttf # replaces the UI font
employees/
default.png # sprite sheet used for all agents
engineer.png # role-specific sheet
designer.png
furniture/
desk_wooden.png # auto-detected as a desk (filename starts with "desk")
bookshelf.png
plant_big.png
This asset pack system provides a straightforward way to visually customize the environment and agents, improving developer experience and clarity when managing multi-agent workflows.
verdict: who should explore outworked
Outworked is a practical tool for developers and AI practitioners who want a hands-on, visual way to orchestrate multiple Claude Code agents with distinct roles and personalities. If you are building complex AI workflows that require task decomposition, multi-agent coordination, and integration with external services, this repo offers a workable architecture and tooling.
The system is macOS-specific and relies on Phaser for the UI, which might limit adoption in cross-platform or web environments. The permission gate system and message bus pattern are solid design choices that make the codebase maintainable and secure, but the complexity of managing many agents and plugins could be a barrier for newcomers.
Overall, Outworked is worth exploring if you want to understand multi-agent orchestration patterns in practice and need a visual, interactive way to manage agent workflows. Its architecture and implementation provide useful insights into combining AI tooling, event-driven communication, and UI visualization to build practical multi-agent systems.
Related Articles
- how awesome-claude-skills turns claude into a real-world action agent — Awesome Claude Skills is a modular Python framework that empowers Claude to perform real-world actions by integrating wi
- Open Design: repurposing coding-agent CLIs into a modular local-first design engine — Open Design turns 12 coding-agent CLIs into a deterministic design engine with 31 composable skills and 72+ design syste
- awesome-copilot: modular community plugins and agentic workflows for GitHub Copilot — awesome-copilot is a community-curated collection of plugins and agents that extend GitHub Copilot with modular, agentic
- Symphony: orchestrating autonomous coding agents with work-level management — Symphony by OpenAI orchestrates autonomous coding agents via work boards and proof-of-work validation, shifting AI codin
→ GitHub Repo: outworked/outworked ⭐ 279 · TypeScript