Hey all, we just released a methodology to develop local, reliable and fast System 1 agents!
What if an AI agent could call tools without writing a single word?
Our new preprint introduces System 1 agents: local AI agents that cannot hallucinate by design, because they never generate text. A deterministic compiler turns each request and the tool’s schema into multiple-choice questions, and a decision model answers them by choosing among options that come only from the request, the schema and the tool catalogue.
We developed six System 1 agents and compared them with nine LLM-based agents on 1,952 requests to 157 medical imaging and biomedical tools from 18 MCP servers:
→ 97.0% correct tool choice, significantly ahead of all local LLM-agent settings
→ Zero invented values in 10,888 calls, against 5.5–19.7% of LLM-agent calls
→ Missing information? It asks the user. LLM agents made it up in up to 62.1% of such requests
→ More accurate and cheaper: 92.9% fully correct at 2.6 s and 0.228 Wh per correct request, against 91.3%, 4.5 s and 0.413 Wh for the best LLM agent
→ Everything runs on one local GPU, so sensitive data never leave the institution
Paper page: [Paper page - System 1 agents: a method for local, reliable and fast agentic AI](https://huggingface.co/papers/2611.21053)
Code and data: [GitHub - eltzanis/SystemOneAgents: A method to develop System 1 Agents · GitHub](https://github.com/eltzanis/SystemOneAgents)
Feedback and questions are very welcome!