Strands Decider 2B: what it is and where it sits in an agent harness

Strands Decider 2B is an open-source decision model from AWS's Strands Labs. It reads a situation and a question, then picks one of the options it was given without writing any text. It came out on 1 October 2026 under Apache-2.0 with 1.9 billion parameters, and AWS measured a median of about 115 ms per question on an RTX 3090. The model itself is not built into the Strands harness. It runs as a local server, and a Strands Agents intervention handler that you write asks it for a verdict before each tool call.

What is Strands Decider 2B?

Strands Decider 2B is a System One model: a small model that answers typed questions with probabilities instead of generated text. A question can be yes or no, one of N options, or a score on a rubric. Strands Labs, the experimental arm of AWS's open-source Strands Agents project, published it on 1 October 2026 in a launch post by Marc Brooker, Mike Chambers and Fabio Nonato de Paula.

It began as Brooker's home project, "Hobson", trained on his own GPU after TypeSafe launched Jev on 15 September (Brooker). The launch checkpoint is hobson-v19. Since then the team has shipped v21 (5 October), the reference checkpoint strands-decider-2B-qwen3.5-v1-2610 (9 October) and four Gemma 4 sizes (changelog), so the name now covers a family. The download is free, and an Amazon spokesperson told VentureBeat there is no hosted API.

Strands Decider scores every option in one forward pass

Strands Decider 2B keeps the torso of Qwen3.5-2B-Base and drops the head that predicts the next token. In its place is a pointer head of about a million parameters that compares the hidden state at the answer position with the hidden state at the end of each option. Training touches only that head and a rank-16 LoRA. A single forward pass returns one probability per option, with no decoding loop.

The Strands Decider 2B architecture, from the launch post. Image: strands-agents/harness-sdk by Strands Labs, Apache-2.0.

The launch post's example routes a support ticket:

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"

It picks billing with a confidence of 0.768 (billing 0.845, retail 0.091, sales 0.064). The first run also downloads the 4.5 GB Qwen base model, since Hugging Face holds only the adapter and the head. The context window is 4,096 tokens, and longer input is cut unless --strict-window is set (inference docs). It runs on CUDA, Apple MPS, MLX or a CPU. There is no official GGUF, and stock llama.cpp cannot run the community one, which needs a patched runtime.

Strands Decider runs beside the agent and answers at a typed checkpoint

Strands Decider 2B runs as its own local HTTP server (POST /v1/systemone on port 8099 in the example), the same shape as a Jev-style sidecar. What Strands adds is the seam: the Strands Agents SDK's intervention API lets a handler pause the agent loop before a tool call, a model call or a whole invocation, and return a typed action that the loop has to obey.

The interventions docs list five actions: Proceed, Deny, Guide (feedback the model retries with), Confirm (pause for a person, tool calls only) and Transform, with "deny > confirm > guide > transform > proceed" as the precedence. The API shipped in Python v1.43.0 on 12 June 2026, before the decider existed.

In the repository's example gate, before_tool_call sends the conversation and the proposed call to the decider. It asks two yes-or-no questions: are the arguments grounded in what the user said, and is the call premature?

answers = self._decider.ask(state, QUESTIONS)
args_grounded = answers["args_grounded"]["noul"]
premature = answers["premature"]["noul"]

gate = ToolCallReviewer(decider)
agent = Agent(
    model="us.anthropic.claude-haiku-4-5-20251001-v1:0",
    tools=[get_weather],
    interventions=[gate],
    system_prompt=SYSTEM_PROMPT,
    callback_handler=None,  # the gate's own log is the interesting output
)

When the grounding answer falls below a hand-picked 0.45, or the prematurity answer reaches it, the gate returns Guide and the Claude Haiku agent asks the user instead of guessing. With Jev or Laya beside an LLM, your own code calls the decider and acts on the answer. In Strands, the gate judges a tool call the model has already chosen, and the loop already knows what each verdict means.

The handler does not care what backend sits behind it: "Strands has no opinion about what goes inside it", says the launch post. Ollaya serves the same model behind a Jev-compatible /v1/systemone endpoint, so it also runs as a plain sidecar outside Strands, as in the guide to running System One models locally.

Which harness steps can a 2B decider gate?

As of 22 October 2026, only one harness step has a Strands Decider example with code: before_tool_call, where the gate checks argument grounding and timing and returns Guide or Proceed. The launch post says "equivalent hooks" exist around the model call and the whole invocation, but shows no decider code for them. Native decision-model support in the Strands SDK is mostly still open pull requests.

The harness's own "smart" intervention setting still uses an LLM risk classifier, not a decision model. No Strands document ties the decider to retrieval, to stopping the loop or to reasoning effort. The post on what a decision model should decide covers which of those calls suit a small model.