Strands Decider is a small, fast decision model for agentic AI. It answers typed questions about a state instead of generating text. Use it when your application already knows the valid outputs and needs a choice, a yes-or-no probability, or a score with a calibrated measure of confidence.
Decider runs on a local CPU, an Apple-silicon Mac, or an NVIDIA GPU. You can call it from the command line, serve it over HTTP, or place it inside a Strands agent.
What Decider returns
Section titled “What Decider returns”Decider supports three question types:
| Type | Use it to | Result |
|---|---|---|
noul | Ask a yes-or-no question | The probability that the answer is true |
choice | Pick one option from a set | The selected option, every option’s probability, and confidence |
score | Rate something on an ordered scale | The expected score, the distribution across levels, and confidence |
One request can ask several questions about the same state. Decider encodes the shared state once, then evaluates each question independently. This keeps multi-question decisions fast without letting one question influence another.
When to use Decider
Section titled “When to use Decider”Use Decider when the output space is known before the request and the same kind of decision happens repeatedly. Common uses inside agentic workflows include:
- routing a request to a model, tool, team, or queue;
- checking tool arguments, grounding, safety, or policy compliance;
- scoring model responses for evaluation;
- handling routine decisions in a hybrid agent while an LLM handles harder, open-ended work.
Use an LLM when the task needs generated text, conversation, coding, summarization, or complex reasoning. Decider classifies and scores; it does not generate a response.
Try one decision
Section titled “Try one decision”Install the package and ask the published 2B model to route a support request:
pip install strands-decider
strands-decider ask StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 \ --state "Help! My payouts have been failing for 3 days." \ --choice "Which team should handle this?=billing,technical,sales"The result names the selected option, reports a probability for every option, and includes a confidence for the choice. See Inference for the complete installation, device, command-line, and HTTP server guide.
Explore Decider
Section titled “Explore Decider”- Inference: install Decider, ask questions, and serve the HTTP API.
- Images: ask questions about image inputs with supported models.
- Architecture: understand the pointer head, question types, calibration, and shared-prefix cache.
- Model names: choose and interpret a published model identifier.
- Strands Agents example: gate a tool call with Decider inside a Strands agent.
New to Decider? Continue with Inference to run the model locally.