Seven AI agents walk into a codebase

What it does

LangManus is a multi-agent automation framework that routes tasks through a hierarchy of specialist models. A supervisor delegates work to agents dedicated to research, coding, browser automation, and reporting, while a coordinator and planner handle upfront routing and strategy. It is fundamentally an orchestration layer—sensible glue code that wires together Tavily search, Jina neural retrieval, a Python execution environment, and browser control into a structured workflow.

The interesting bit

The framework uses a three-tier LLM setup with separate configurations for reasoning, basic tasks, and vision-language work, letting you route jobs to the appropriate model tier instead of leaning on a single endpoint for everything. Each agent’s behavior is defined in plain markdown files under src/prompts/, making the system’s logic surprisingly legible without reading source code.

Key highlights

• Seven-agent hierarchy: Coordinator, Planner, Supervisor, Researcher, Coder, Browser, and Reporter
• Three-tier LLM configuration to match model capability with task complexity
• Agent prompts stored as markdown templates, editable without touching code
• FastAPI server with Server-Sent Events streaming for real-time responses
• OpenAI-compatible API and explicit support for open-source models like Qwen

Caveats

• Requires multiple external API keys (Tavily, optional Jina, and separate keys for each LLM tier)
• Browser automation depends on a local Chrome instance
• Default web UI lives in a separate repository

Verdict

Worth a look if you want a pre-built, opinionated multi-agent scaffold for research and coding workflows. Skip it if you need a fully self-contained, offline-capable system or are allergic to API key management.