ADK Agent Playground

Project · 2025

Most multi-agent demos are single-purpose toys. I wanted to build enough of them in one framework to see the patterns: where hierarchical delegation works, where sequential pipelines are better, and where agents waste tokens arguing with themselves. Six systems, one shared infrastructure, any LLM via OpenRouter.

Each agent tackles a different domain – deep research, code review, structured debate, market analysis, content summarization, and iterative writing – using hierarchical delegation and sequential pipeline patterns. All run through OpenRouter, so any LLM can power them: Claude, Gemini, DeepSeek, Llama, or any other model. Switching models is a one-line change.

The hardest design problem wasn't the multi-agent orchestration. It was deciding when an agent should stop researching and start writing. The shared infrastructure layer handles model configuration, rate limiting, caching, cost tracking, and session management. Each agent is self-contained in its own directory with a standard structure (root agent, sub-agents, tools).

Agents

Architecture

The Deep Research agent is the flagship. It uses hierarchical delegation: an orchestrator plans the research, dispatches specialist sub-agents in parallel, then synthesizes a cited report.

Shared Infrastructure

The shared layer provides reusable components across all agents: model configuration via OpenRouter and LiteLLM (swap models with one line), an agent runner and CLI helpers for consistent execution, rate limiting with retry logic and HTTP connection pooling, cost tracking per agent run, session and state management, and a tool registry for external integrations.