The Pulse
Strands Agents Releases a 2B Model for Agent Decisions
Strands Agents released Decider 2B, an open model designed to make fast yes-or-no and option-selection decisions inside AI agent workflows. The model can run locally, but its developers say it is not suited to complex reasoning, chatbots or
AI.info Team · Written with AI ·
Strands Agents is betting that some agent decisions do not need a text-generating model. On October 1, the team released Strands Decider 2B, an open model built to choose among options or score a prompt, not to write answers. The trade-off is explicit: the smaller model is meant to handle quick, bounded choices, while general-purpose language models retain the harder reasoning and generation work.
A decision model, not a smaller chatbot
Decider’s job is to answer questions such as whether a proposed tool argument is grounded in a user’s message, which team should receive a request, or which model should handle a task. Unlike a conventional language model, it selects from choices or assigns scores rather than generating arbitrary text. Strands says that design can return a decision with a confidence score and answer multiple questions about the same input efficiently.
The model starts with the Qwen3.5-2B language-model torso, but removes its text-generation head and replaces it with a pointer head that scores candidate answers. Strands says the replacement head has just over one million parameters and that the torso is fine-tuned with a rank-16 LoRA adapter. The result is meant to make a decision in one pass rather than generate a stream of text.
“The flip side is that this approach (generating all outputs in a single parallel pass) makes it significantly worse at solving complex problems than reasoning models.”
Marc Brooker, Mike Chambers and Fabio Nonato de Paula, authors of the Strands Agents announcement
Speed claims come with a benchmark caveat
Strands reports a median decision time of about 115 milliseconds on a local Nvidia RTX 3090, with latency rising approximately linearly as task size grows. The team reports a median of about 153 milliseconds for small tasks on an M3 MacBook. Those are the company’s measurements, not a guarantee for every device or workload.
On JevBench’s public set, Strands says Decider ranked third among 33 models in the 2B class, and first among 30 after models just over 2 billion parameters were excluded. The team also says it answered all of JevBench’s easy tasks correctly. Those results describe a narrow benchmark and do not establish that the model can replace a larger system on open-ended agent work.
The point is to gate an agent’s next move
Strands identifies model routing, tool selection, guardrails, evaluations, memory and context management as possible uses. Its worked example pairs a local Strands agent with a local Decider model: when a user asks for the weather without naming a city, the larger agent proposes a tool call using a guessed location. Decider checks whether the proposed arguments came from the conversation and whether the call should wait for clarification.
In that example, the intervention stops the unsupported call and sends the agent back to ask which city the user meant. The post stresses that the sample’s questions, thresholds and policy were chosen by hand; it presents the workflow as an illustration, not a recommendation. Developers can install the strands-decider command-line tool and access the code repository and model weights.
Complex work stays with larger models
Decider is designed as a supporting component, not a standalone assistant. Its restricted output makes it unsuitable for coding, chat, summarization and problems that require multi-step reasoning, according to the Strands announcement. The proposed division of labor is concrete: use a larger language model for difficult or open-ended tasks, and let Decider handle bounded checks such as whether a tool call is grounded before it runs.