# Strands Decider

> AWS's small open decision model that picks, answers yes/no or rates with a calibrated confidence, locally

- Category: App Orchestration / Structured Output
- Language: Python
- Licence: Apache-2.0
- Repository: https://github.com/strands-labs/strands-decider
- Website: https://strandsagents.com/blog/introducing-strands-decider/
- Canonical page: https://ai-tldr.dev/tools/strands-decider/
- Updated: 2026-10-01

## In short

Strands Decider is a "system one" decision model for agentic workflows from AWS Strands Labs, the experimental arm of the Strands Agents project. It does not generate text. It sits in the Structured Output category of the App Orchestration tools and is open-source, written in Python and licensed Apache-2.0. To get started: install the package.

## Overview

Strands Decider is a "system one" decision model for agentic workflows from AWS Strands Labs, the experimental arm of the Strands Agents project. It does not generate text. You give it a piece of state and one or more questions — pick one of these options (`choice`), yes or no (`noul`), or rate on an ordered scale (`score`) — and it returns an answer with a calibrated confidence for every option.

The reference model, `strands-decider-2B-hobson-v19`, starts from the Qwen3.5-2B base model. The language-model head is removed and replaced with a pointer head of about one million parameters that scores each supplied option by comparing hidden states, plus a rank-16 LoRA adapter. One forward pass answers the question, with no decoding loop, so every output is one of your options and nothing has to be parsed.

The README reports 72.3% (167 of 231) on the public JevBench set with a Brier score of 0.342 and an ECE of 0.052, a 115 ms median latency per question on an RTX 3090 and a 153 ms warm median on an M3 Pro Mac. The code, training setup and data source list are Apache-2.0, and a full training run takes about 11 hours on one RTX 3090. Within structured output it sits beside Jeff, Laya and Jeeves, which answer the same Jev-style questions.

## What it does

- Three question types — `choice`, `noul` (yes/no) and `score` — and several can be asked about the same state in one call
- A calibrated confidence on every answer; the README reports about 95% accuracy on unseen tasks when confidence is 0.9 or higher
- Single forward pass through Qwen3.5-2B with a pointer head, so outputs are always one of the supplied options
- `strands-decider ask` CLI and a `strands-decider serve` HTTP server with a `/v1/systemone` endpoint
- Runs on NVIDIA GPUs, on Apple silicon with optional MLX acceleration, and on CPU
- Training scripts included: about 11 hours on one 24 GB RTX 3090 or about 1 hour 10 minutes on eight H100s

## Getting started

Strands Decider installs from PyPI and downloads the reference model from Hugging Face on first use.

1. Install the package

```bash
pip install strands-decider
```

2. Ask a choice question

   The answer comes back with a probability for each option.

```bash
strands-decider ask StrandsAgents/strands-decider-2B-hobson-v19 \
  --state "Help! My payouts have been failing for 3 days!" \
  --choice "Which team should handle this?=billing,sales,retail"
```

3. Run it as a server

```bash
strands-decider serve StrandsAgents/strands-decider-2B-hobson-v19 --port 8000
```

4. Send a yes/no question over HTTP

```bash
curl -s localhost:8000/v1/systemone \
  -H 'content-type: application/json' \
  -d '{
    "state": "Help! My payouts have been failing for 3 days!",
    "questions": {
      "is_urgent": {"type": "noul", "instructions": "Does this convey urgency?"}
    }
  }'
```

## When to use it

- Routing a request to the right model or team before a larger LLM is called
- Choosing a tool and checking its arguments before an agent runs it
- Guardrails and policy classification with a confidence threshold
- Scoring whether an agent's output is good enough, as a cheap local evaluator

## Version history

- 2026-10-01 2B hobson-v19: AWS Strands Labs released Strands Decider 2B: an Apache-2.0 decision model on Qwen3.5-2B with a pointer head that answers choice, yes/no and score questions with calibrated confidence in about 115 ms on an RTX 3090.
