Purpose
Agr is a decision model released by Command Code. You send it a state (text or JSON) and one or more typed questions, and it returns typed answers with probabilities. It is suitable for decision scenarios with a fixed set of answers, such as routing tickets, selecting tools, and approving actions.
Main Capabilities
- Directly scores a fixed set of options and returns a probability for each option, rather than generating text that needs to be parsed.
- Answers all questions for the same state in a single forward pass, independently of one another; one question cannot attend to another question.
- Three question types, namely the System One primitives:
- choice: choose one from several named options; criteria maps each name to a description.
- score: evaluate the state according to the levels listed in order in criteria.
- noul: yes/no judgment.
- confidence uses the System One definition, not the probability of the highest option.
- Each question contains two fields, type and instructions, and the input is limited to text.
Versions and Evaluation
- Agr is available in two sizes: Agr and Agr-flash.
- Agr scores 58.15 on the 38 benchmark items counted in Decision Index 0.2.1.
- Evaluation runs on all 150,759 requests in the public Decision Index suite, revision 87d4650, with no truncation. Jev, Kev, and Rune scores are taken from the public leaderboard as of September 28, 2026.
- The index averages random-corrected skill across five domains; the numbers are skill × 100, not raw accuracy or F1. For all domains, higher is better, including ForecastBench (the index converts Brier scores into skill scores). 3% of BRIGHT requests exceed Agr's 32,768 token limit and are counted as errors.
- Agr-flash is not used for safety decisions.
- Each repository contains the weights, tokenizer, and configuration.
Usage
The environment requires Python 3.11 or newer. Installation also pulls in PyTorch and the tested Transformers version. The first run downloads the weights from Hugging Face: Agr's bf16 weights are 61.43 GB, the official service is provided on a 96 GB GPU; for Agr-flash, use CommandCode/agr-flash.
The server receives System One requests at /v1/systemone, and can be called with curl, Python, or TypeScript.
TypeSafe Python SDK:
TypeSafe JavaScript SDK (Node.js 20 or newer):
The server does not require a key by default, so the API key can be any string. If --api-key or AGR_API_KEY was provided at startup, then that key must be passed in.
Evaluating on your own data: reports accuracy, Brier score, calibration error, and latency. For the label format and agr compare, see DESIGN.md.
Use cases
- Classification-style decisions with a fixed set of answer options, such as ticket routing, tool selection, and action approval.
- Situations where you need to directly obtain a probability distribution over the options, rather than free text.
- Situations where you want to pose multiple mutually independent questions in parallel on the same state.
Limitations
- Only English text is supported; images are not supported.
- Answers are limited to the options you provide; no text or explanation is generated.
- Before relying on probabilities and thresholds, check them first on your own labeled examples.
- Changing the order of the options may change the answer.
- Neither model is an independent safety filter, nor is it suitable as the sole basis for major decisions involving people.
Acknowledgments and license
Agr is built on TypeSafe's Jev and its System One API, as well as Archer Hume's analysis of the Jev architecture; it also references Jared Palmer's Kev, Mark Marosi's Decider (answer-slot prompt layout and temperature rules), and Invergent AI's Surogate (option order averaging). Thanks to apolinário and the Decision Index maintainers, Google DeepMind for Gemma, Hugging Face for SmolLM2, and the authors of the datasets used. Agr is not affiliated with TypeSafe or the systems compared in this document.
Agr is released under Apache 2.0. Agr uses Google DeepMind's Gemma 4 31B IT; Agr-flash uses Hugging Face's SmolLM2-360M. Both base models are licensed under Apache 2.0. For attribution, see NOTICE.md.
Not yet provided
Apart from the above, details such as charges, pricing, platform support scope, and specific usage steps are not yet provided.
