Experiment Automation Agents

EAA is a runtime and harness for agents and workflows, designed for running scientific experiments especially for those needing vision. EAA turns a multimodal LLM into an experiment operator with features including the following: - A built-in LLM chat loop - An image-followup mechanism allowing LLMs to "see" images yielded by tools, even if the LLM provider's API doesn't allow images in tool responses - Automatic image captioning - Support of MCP servers and external HTTP endpoints as tools - Support of agent skills - Built-in tools including - File read/write - Python/bash coding (including Bubblewrap sandboxing) - Local environment setup with uv - Image rendering tool allowing LLM to see floating point TIFFs - Subagent and sub-task manager spawning - Fully customizable "task manager" data structure - A WebUI for easy interaction and visualization of sessions

EAA is not only for LLM agents! Fully logic-driven or rule-based workflows - such as Bayesian optimization - can be packaged as task managers. Use MCP servers as RPC calls through the built-in MCP-RPC wrapper, and display status and progress using the same WebUI. Logic-driven task managers can be added to the sub-task manager tool so that an LLM agent can launch them.

Start with the quick start, or use the guide for installation, configuration, tools, skills, memory, and WebUI usage.