How h2oGPTe Agents Work: Multi-Model, Code-First
- Agentic AI
- Enterprise h2oGPTe
This lesson explains what makes h2oGPTe's agentic AI architecture different from typical multi-agent frameworks. It covers three design choices: routing each task to the right specialized model, such as Whisper for transcription or Sonnet 3.5 for reasoning, using a single flexible agent with modular tools instead of many purpose built agents, and having the agent write and execute code directly through the CodeAct framework. It then walks through a live example: the agent analyzes Tesla's stock price, switches to a backup data source when its first API fails, and scrapes a web search for sentiment context.
Key takeaways
- How h2oGPTe routes each task to a specialized model, using Whisper for audio, Mississippi or DocTR for documents, and Sonnet 3.5 for reasoning
- Why h2oGPTe uses a single flexible agent with modular tools instead of multiple purpose built agents and rigid predefined workflows
- What the CodeAct framework does: the agent writes, runs, and refines its own code as its main way of taking action
- How the agent handled a failed stock price API by automatically switching to an alternative data source without manual intervention
- The steps the agent took to enrich its stock analysis: running a web search, then scraping and structuring the results
Start owning your AI