Experiments and recipes for AI agents: multi-agent systems, code-execution sandboxes, and an agentic system for playing Factorio/Polytopia.
| Area | Description |
|---|---|
| Multi-agent systems (Google ADK) | Examples built with Google's Agent Development Kit — sub-agents, sequence agents, loop agents — using LiteLLM for inference. |
| Coding agents + sandbox | Agent harness that reformulates questions as coding problems, generates code (e.g. Qwen3-Coder), and runs it in a sandboxed environment with resource limits. |
| Factorio | Agentic system that plays Factorio. See the Factorio README for setup. |
The coding-agents pipeline runs an LLM-backed codegen service that sends code to a sandbox for execution. Example: "How many r in strawberry?" is turned into code, executed, and the result is returned.
For architecture and data flow (including the full "how many r in strawberry?" flow), see Coding Agents README.