通常のチャットでAIエージェントのオーケストラを実行する方法

コードエディターで1つのAIアシスタントを使っているだけの人から、複数の自律型エージェントに移行すると、すぐに制御の問題に直面します。エージェントはバックグラウンドプロセスを生成し、トークンを消費し、何か問題が発生すると、Dockerの奥底のどこかで静かに環境を壊し始めます。APIキーやGitHubへのアクセス権を付与する必要がある場合は、さらに深刻です。自律的にコードを生成・実行するスクリプトに個人トークンを渡すのは、控えめに言っても緊張するものです。

最近、AgentScopeチームのAgentTeamsプロジェクトを見つけました。Matrixプロトコル 기반으로ニューラルネットワークエージェントを統合チーム環境にパッケージするランタイムプラットフォームです。抽象的なバックグラウンドデーモンではなく、通常のMessengerであなた、マネージャーエージェント、 Specialized worker agents all participate in the same chat, enabling direct communication and coordination between team members.

The architecture leverages a Manager-Workers model where the manager decomposes high-level requests into subtasks and distributes them to specific workers, who then execute those tasks in isolated environments. The Matrix messenger acts as the unified communication bus for all interactions, ensuring that every agent-to-agent conversation is visible and traceable throughout the system.

This approach provides complete transparency—if a worker deviates from its assigned task, you can simply message it directly in the chat to correct the course without restarting the entire workflow. The main security concern with autonomous agents is that they typically need direct access to API keys and tokens to perform actions like creating Pull Requests or calling external APIs, which creates vulnerabilities in the system architecture.

AgentTeams addresses this through a built-in Higress AI Gateway that acts as an intermediary. Workers only receive temporary consumer tokens, and all external requests—whether to LLM APIs, GitHub, or MCP servers—must route through this gateway. The real credentials stay locked inside the gateway and never reach the worker containers, so even if malicious code tries to extract environment variables, it only gets the local gateway access token.

The platform also supports combining different agent types in a single room, with three runtimes available: OpenClaw on Node.js for tool calling and task coordination, QwenPaw on Python for lightweight tasks and browser automation, and Hermes for autonomous code writing with an isolated terminal sandbox. This flexibility lets you assign a deterministic OpenClaw agent as the group leader while connecting Hermes for code writing and debugging. Agents communicate through built-in mentions and share a MinIO file store, which significantly reduces context window usage since they reference files in storage rather than passing large code chunks or logs directly in messages.

Architecturally, AgentTeams is delivered as interconnected services with a Controller managing resource lifecycle and declarative configuration, a Higress AI Gateway handling traffic and authorization, Tuwunel and Element Web providing Matrix server and web client functionality, and MinIO storing artifacts generated by agents. Docker is sufficient for local mode deployment with minimum requirements of 2 CPU cores and 4 GB of RAM, though 8 GB or more is recommended if running 3-4 workers simultaneously. Installation launches with a single command that prompts for an LLM provider key (any OpenAI-compatible endpoint works), configures local services, and provides a link to Element Web.

For production environments, the project includes an officially supported Helm chart for Kubernetes that allows defining workers and teams through Custom Resource Definitions. Helm automatically runs a preflight check before launching the controller, sending a test request to verify the API key is valid and limits aren't exhausted.

The project stands out from typical console scripts by using a familiar messenger interface instead of another web dashboard. This approach is surprisingly practical since Matrix can be accessed from mobile clients like FluffyChat or Element Mobile. However, the hardware demands are substantial—running multiple containers alongside MinIO and the gateway consumes memory quickly, making it excessive for simple IDE assistance tasks. For teams needing a complete pipeline of AI developers without compromising API key security or losing operational visibility, this is worth exploring.