Papers
arxiv:2611.21053

System 1 agents: a method for local, reliable and fast agentic AI

Published on Nov 16
· Submitted by
Eleftherios Tzanis
on Nov 20

Abstract

A local System 1 agent framework compiles tool schemas and requests into multiple-choice questions answered by a one-pass decision model, enabling accurate biomedical tool calls without text generation or invented content.

Agentic artificial intelligence could automate analyses of sensitive clinical data, but current agents depend on large language models (LLMs). Cloud LLMs expose patient data. Local LLMs hallucinate, and an invented value in a tool call propagates through every later step. Here we present a method to build System 1 agents: local agents that call tools without generating text. A deterministic compiler turns each tool's schema and the request into multiple-choice questions whose options come from the request and the schema, then a one-pass decision model answers them. On 1,952 requests to 157 biomedical tools, the best System 1 agent chose the right tool, or correctly declined, for 97.0% of requests, significantly ahead of all 18 local LLM-agent settings (best 95.2%). The System 1 agent filled 97.0% of 6,056 stated argument values. No System 1 call contained invented content, against 5.5-19.7% of LLM calls. The System 1 Agents framework is freely available as open-source software at https://github.com/eltzanis/SystemOneAgents.

Community

Turning each tool call into small multiple-choice questions, so the agent never has to invent a value, is a neat way to deal with hallucination in clinical settings. Running it locally also keeps patient data off a cloud model. How does it handle a request where the right answer is not among the options the compiler builds from the schema?

·

Thanks! Every question has an exit option: 'none of these' for the tool and 'not specified' for each argument value. If the right value is not among the candidates, the model is expected to pick that option. For a required argument, the agent then leaves it empty and asks the user with a fixed prompt built from the schema. For an optional one, it keeps the tool's default, and if no tool fits, it declines.

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

The part I'd actually deploy is the compiler, not the one-pass decider — turning a schema plus a request into a multiple-choice question moves the failure from "the model invented a string" to "the model confidently chose the wrong enum option," and the second one is far easier to log and bisect. What I can't work out from the method is what happens to arguments that can't be enumerated: file paths, free text, an ID handed back by the previous tool call. Every tool-calling bug I've shipped in the last year lived in exactly that gap. The 157 bio tools also read like hand-curated, stable schemas, and mine drift weekly — someone renames a parameter and the compiler is quietly reasoning over stale options. So: does the compiler re-derive per request, and what does that cost on a single local GPU where the speed claim has to actually hold? If it's amortized against a cached schema that's a fine design, but it also means "fast" is downstream of how disciplined your tool registry is, which is the part nobody wants to own in prod.

·

Thanks, this is a sharp read, and I agree with the framing: a failure becomes a logged choice with a probability. s1a plan prints every question, the top options and their probabilities, and every run can be replayed from the stored answers.

Non-enumerable arguments: paths, free text and IDs are not enumerated from the schema. They are filled from candidates the compiler extracts, plus a "not specified" option. Candidates come from four sources: the request (paths, numbers, quoted strings, identifiers and compact codes, short phrases for name and search arguments, text after a label), forms derived by fixed rules (unit conversions, paths composed from written parts), values the schema lists, and the environment and earlier steps (the header of a CSV that another argument points to, and earlier tool results). So an ID handed back by the previous call is a candidate. The compiler can extract or derive text but never compose new text, so an argument that needs freshly written prose is out of reach.

Schema drift: the questions are compiled from the chosen tool's schema on every request, and the compiler has no rules for particular tools; gaps were fixed by general rules about how requests write values. The schemas are each server's tools/list output, unchanged, including third-party servers. The catalogue is a JSON snapshot so that tool selection never has to start a server, and s1a catalog rebuilds it from your servers in one command. A renamed parameter is picked up as soon as the snapshot is refreshed, with no retraining or code change. But you are right that a stale snapshot means stale options: the agent does not watch for tool-list changes today, so keeping the registry fresh is on you, as with any agent that caches tool definitions.

Cost: compiling is CPU-side text processing; the expensive part is the roughly 12 decision-model questions per request, and independent questions are sent together. The reported times are wall-clock per request (1.2–10.1 s across the System 1 agents), so they include compilation, but we did not break out the compiler's share.

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2611.21053
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2611.21053 in a model README.md to link it from this page.

Datasets citing this paper 1

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2611.21053 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.