It comes from Strands Labs, the experimental arm of Strands Agents, and rides the current wave of "system one" decision models: small models that answer the fast, rote questions so your big LLM can spend its tokens on the hard ones. Under the hood it's a Qwen3.5-2B-Base torso with the language-modeling head thrown away, replaced by a ~1-million-parameter pointer head that scores each option against the model's answer position. One forward pass, no decoding loop; the torso is adapted with a rank-16 LoRA adapter and the head runs in fp32. Because the head carries no per-option parameters, your label sets are defined by the request — never baked into the weights.

What you’ll need

  • Python 3 and pip. pip install strands-decider is the whole install. The 1.9B checkpoint runs on an Apple-silicon Mac (MPS, or --device mlx for 1.4–1.6× faster than MPS), on CPU, or on a CUDA GPU — the 115 ms figure is the RTX 3090 reference.
  • A few GB of disk for the 1.9B-parameter checkpoint, pulled from Hugging Face on first use.
  • No training, no labels. You ask questions in plain text; the model answers them. That is the entire point.

1. Install and ask your first choice question

Install the CLI and triage a support ticket — a classic choice question, straight from the project's README:

pip install strands-decider
strands-decider ask StrandsAgents/strands-decider-2B-hobson-v21 \
  --state "Help! My payouts have been failing for 3 days! " \
  --choice "Which team should handle this?=billing,sales,retail"

Example output from the project:

choice_0 -> billing (confidence 0.835)

  billing                  0.890
  sales                    0.056
  retail                   0.054

You get the pick, a confidence number, and a distribution over every option. That confidence is the feature: on short classification tasks it has never seen, answers at a confidence of 0.9 or more are right about 95% of the time (evaluation/results.md in the repo). Below that, the README says to confirm or ask a person. In an agent, this maps directly to a policy: act above your threshold, escalate below it.

2. Ask a yes/no question with noul

The second question type is a yes/no — called noul, measuring how strongly the model leans "yes":

strands-decider ask StrandsAgents/strands-decider-2B-hobson-v21 \
  --state "Help! My payouts have been failing for 3 days! " \
  --noul "Does this convey urgency?"
noul_0 noul = 0.875

Closer to 1 leans more toward yes. This is your guardrail primitive: policy classification, grounding checks, "should a human see this before it goes out."

3. Score on an ordered rubric

The third type rates things on a scale — useful for evaluations and prioritization:

strands-decider ask StrandsAgents/strands-decider-2B-hobson-v21 \
  --state "Help! My payouts have been failing for 3 days! " \
  --score "How frustrated is the writer?=calm,frustrated,depressed"
score_0 score = 1.07 (confidence 0.602)

  0: calm                                     0.140
  1: frustrated                               0.648
  2: depressed                                0.212

4. Batch all three in one call

The state is read once and each question adds only its own tokens, so combining questions is cheap:

strands-decider ask StrandsAgents/strands-decider-2B-hobson-v21 \
  --state "Help! My payouts have been failing for 3 days! " \
  --choice "Which team should handle this?=billing,sales,retail" \
  --noul "Does this convey urgency?" \
  --score "How frustrated is the writer?=calm,frustrated,depressed"

You now have routing, urgency, and frustration in a single batched call — the three things a triage loop needs before it decides what to do.

Terminal showing a decision model's probability bars with a calibrated confidence score
A decision with a confidence you can act on: the distribution over options plus a calibrated score. Illustration generated for this article.

5. Serve it over HTTP for your agent

For real wiring, run it as a server — the README's own pattern for putting decisions inside agentic workflows:

strands-decider serve StrandsAgents/strands-decider-2B-hobson-v21 --port 8000
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?"}
    }
  }'

The response carries the model name, the typed answers, token usage (note: one output token — nothing was generated), and the latency in milliseconds. From here the README's suggested uses read like an agentic-systems checklist: model routing (pick the right LLM for a task), tool selection (which tool the agent should call next), argument checking (verify a tool call before it runs), triage, guardrails, evaluations, and hybrid agents — the big LLM makes the hard decisions, the decider makes the rote ones. The repo ships a worked examples/strands/ example that gates a tool call with a before_tool_call intervention.

A tiny fast neural network racing ahead of a huge slow language model with a stopwatch
One forward pass, no decoding loop: the decider answers while the LLM is still warming up. Illustration generated for this article.

What you built

A ticket-triage decision pipeline with five moving parts: an install, a choice question that routes a ticket to billing with 0.89 of the distribution mass, a yes/no urgency check, a frustration score, all three batched into one call, and an HTTP endpoint your agent can query at ~115 ms a question. No JSON parsing, no prompt-format fragility, and a confidence number you can turn into an escalation policy.

Honest limitations

  • It cannot explain itself. A decision model never generates text, so you get the pick and the confidence but no rationale. If a decision needs a paper trail, pair it with an LLM — the hybrid-agents pattern is in the README for exactly this reason.
  • Calibration holds where it was measured. The 95%-at-0.9 figure is for short classification tasks the model has never seen; on long, messy, or domain-specific input, recalibrate before trusting the number. On JevBench v1 the current v21 scores 0.762 accuracy (176 of 231) with expected calibration error 0.064 — and the hard tier sits at 0.550. Hard questions still beat it.
  • Read the README's own caveats. The v19/v21 comparison columns are two single runs, not a measured gain; run-to-run standard deviation is 3.2 tasks, so treat any gap under about 10 tasks as unresolved.
  • It's early. Sixteen commits from an experimental lab (Strands Labs, created September 29, 2026) — expect the API to move. The vision variant (--vision) needs the [vision] extra and transformers 5.18 or later; I left it out of this walkthrough deliberately.

Still: for the rote decisions inside every agent — routing, tool selection, guardrails — a 2B model that answers in 115 ms with a confidence you can budget against is the cheapest upgrade on the board.