System 1 Agents: a local agent method that generates no text is released
Understand this story
Eleftherios Tzanis and Michail E. Klontzas released a paper and code repository proposing the System 1 Agents method: a deterministic compiler turns tool JSON schemas and user requests into finite multiple-choice questions, a local decision model gives the probability of each option in a single forward pass, and code then assembles the MCP tool call, with no text generated at any point. The authors describe the method as aimed at local, reliable, fast agent applications. On runtime requirements, the authors say a CPU suffices, but it must be paired with llama.cpp's llama-server and an NVIDIA GPU able to hold the decision model (decider-4b starts at 7.2 GiB). Model licenses must be checked per model card: OpenJev is CC-BY-NC-4.0 (non-commercial), and the Perplexity models may require accepting their terms on Hugging Face. The release does not include an MCP server; the tool catalog is only a JSON snapshot of its tool list, and the models are not distributed with the repository.
AI overview based on source reports · Updated 1 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AISystem 1 Agents – a method for local, reliable and fast agentic AI
The authors present System 1 agents, which call tools without generating text: a deterministic compiler turns a tool's JSON schema and the user's request into typed multiple-choice questions, and a local decision model answers each with option probabilities in one forward pass, while code assembles the call. Across 157 MCP tools from 18 medical-imaging and biomedical servers, the best System 1 agent made 93.9% of requests fully correct versus 91.3% for the best local LLM agent, with no invented tool, argument or value detected in 10,888 System 1 calls and 86.2–93.7% correct handling of underspecified requests versus 0.0–5.7% for LLM agents. The repository releases code, data, results and figures, with instructions to reproduce the paper's results on CPU and to build agents for other MCP tool catalogues.
Activity for this story
Not enough continuous observations to show a trend yet.