I'm Nitin Kumar, Lead Project Coordinator at Debugged Pro. I work with software and IT agencies in India and the USA that ship AI agent workflows for clients. Strands Decider landed on October 1, 2026 as Amazon's open decision model for those workflows.

TechCrunch and VentureBeat covered the release the same day. The official brief from Strands Agents fills in architecture, latency charts, and a tool-call demo.

Decision models answer a narrow question from a closed set of options. They do not write paragraphs. For agencies, that means cheaper checkpoints around tool calls, routing, and policy checks while a larger model still handles generation.

Strands Decider open source AI decision model announcement graphic — Feature image: Strands Decider announcement art from the Strands Agents blog. Credit: Strands Agents / AWS Strands Labs.

What Strands Decider changes for agent delivery

Most agent stacks still ask a frontier LLM to decide the next step in prose. That works. It is also slow and expensive when the question is already closed.

Strands Decider sits in that gap. Give it state and a typed question. Get a choice and a confidence score in one forward pass. AWS built it inside Strands Labs, the experimental agent toolkit group at Amazon.

Amazon distinguished engineer Marc Brooker told TechCrunch the spark was TypeSafe's Jev. He wanted a workflow step that answers "what is the next thing for me to do here?" with better reliability from confidence scores and a closed answer set.

Brooker also said AWS customers asked for this. Not every agent step needs a full LLM. Some steps need a fast, structured pick.

How the Strands Decider architecture works

The design is easy to explain in a client workshop. Start with a pretrained LLM torso. Strip the text-generation head. Add a small scoring head.

Base torso: Qwen3.5-2B, about two billion parameters.
Pointer head: scores supplied options against a decision token; just over one million parameters.
Adapter: rank-16 LoRA fine-tune on the torso.
License: Apache 2.0; weights on Hugging Face; code, data, and scripts on GitHub.

AWS calls this pointer design Hobson in materials shared with VentureBeat. The official blog says the released checkpoint is v19 after many iterations. A slot-head earlier design performed worse. Version 20 experiments did not replace v19 as the reference model.

Because the model scores options in parallel, it always returns one of the allowed answers. That helps when your product code expects an enum, not free text. The tradeoff is clear: it is weak at complex open problems and cannot write code or summaries.

Strands Decider 2B architecture diagram with pointer head over Qwen torso — Body image: Strands Decider 2B architecture diagram from the Strands Agents announcement. Credit: Strands Agents / AWS Strands Labs.

Speed, accuracy, and what the charts actually show

Agencies should quote the published numbers carefully. Different charts cover different checkpoints and hardware.

Latency figures you can put in a brief

Official Strands blog: median around 115 ms on a local Nvidia RTX 3090.
Same source: small tasks around 153 ms median on an M3 MacBook.
VentureBeat cites an earlier v18 chart: 106 ms median and 296 ms p95 across 230 requests on an RTX 3090, including HTTP round trip.
Latency grows roughly linearly with task token size.

These are local measurements. They are not a multi-region hosted SLA. Warm-up can spike the first request. AWS excluded a 7.7-second first request from one chart for that reason.

Accuracy and calibration on JevBench

AWS tracks accuracy and Brier score on the public JevBench set. Lower Brier is better for calibration. VentureBeat reads the plotted v19 point at roughly 72% accuracy and a 0.35 Brier score.

The official blog says Strands Decider ranks third of 33 in the 2B class on that public set, and first of 30 if you exclude models just over 2B. AWS also says it ranks first among peers that publish a full training recipe.

Important caveat from VentureBeat: the plotted comparison does not include TypeSafe Jev itself. Do not claim Strands beats Jev on accuracy. Mapika's similarly sized decider looked slightly ahead of Strands v19 on the same chart.

Strands Decider accuracy and Brier score trajectory across training versions — Body image: accuracy and calibration trajectory chart from the Strands Decider announcement. Credit: Strands Agents / AWS Strands Labs.

Strands Decider versus hosted Jev for agency stacks

TypeSafe launched Jev on September 15, 2026 as a System One style decision model. It returns choices and probabilities without generating prose. Jev stays API-only at about $0.042 per million input tokens, with no output-token fee.

Strands Decider takes the opposite product bet: open weights, self-host, full recipe. AWS is not offering a hosted Decider API yet.

Sticker price: Strands download is free; Jev bills per use at a very low rate.
Operating cost: Strands still needs CPUs, GPUs, or cloud instances. AWS gave no per-request cost model.
Control: Strands can stay inside a client VPC or on a laptop for experiments.
Ops load: Jev is call-and-pay; Strands needs deploy, monitor, and upgrade work.
Ecosystem fit: Strands hooks into the Strands agent intervention path for before-tool checks.

VentureBeat's fair summary: Strands is a competitive open alternative, not a proven Jev-killer on benchmarks. The strategic win is reproducibility and local control.

Where agencies should use an AI decision model

The Strands team lists early use cases that map cleanly to delivery work.

Model routing between cheap and expensive LLMs
Tool selection and tool-call allow or deny
Evaluations and offline grading of agent outputs
Guardrails and policy classification
Memory and context management choices
Hybrid agents that keep hard reasoning on an LLM

The weather demo is the clearest teaching example. An eager agent guesses a city when the user asks for weather without a location. Before the tool runs, Strands Decider answers two yes/no questions: are the arguments grounded in what the user said, and is the call premature? Application code then asks for the city instead of inventing one.

That pattern uses Strands' InterventionHandler. Return types include Proceed, Deny, Confirm, or Guide. The questions and thresholds in the demo were chosen by hand. Treat them as a pattern, not a finished safety product.

Risks agencies must put in the SOW

Decision models are not magic security walls. VentureBeat has already covered how adversarial text can steer probabilistic guards. Tell clients that.

Not a security boundary alone. Pair with allowlists, least privilege, and human confirm for high-impact actions.
Calibration can drift on client jargon, languages, and unusual cases outside JevBench.
Thresholds are product choices. Document who owns them and how you retest after model updates.
Self-host cost is real. Free weights do not mean free GPU hours or free on-call.
No text generation. Keep coding, chat, and long reasoning on a generative model.

Diogo Almeida, TypeSafe CEO, told TechCrunch that building useful intelligence here is harder than cloning the architecture. That is a useful reminder for sales decks. Do not promise production-ready judgment from a weekend fine-tune.

Practical rollout for India and USA delivery teams

Here is a three-week shape we use when a client wants cheaper agent checkpoints without rewriting the whole stack.

Week 1: pick closed decisions

List the top agent branches that already map to enums: which queue, which tool, proceed or ask, which model tier. Capture traces. Mark decisions that fire hundreds of times per day. Those are the first candidates for Strands Decider.

Week 2: local spike and threshold draft

Install the open model on a GPU box or a strong Mac for the spike. Replay anonymized traces. Log choice, confidence, and human label. Draft thresholds for Confirm versus Proceed. Keep the larger LLM for generation and hard cases.

Week 3: gate, audit, and client review

Put the checkpoint behind feature flags. Log every decision with input hash, options, scores, and action taken. Review false allows and false denies with the client. For India delivery centers serving USA clients, keep the decision service in the region the contract requires.

If your team needs help wiring checkpoints into a client-ready delivery process, our software development consulting practice covers agent architecture reviews, evaluation harnesses, and runbooks. Related reading on our blog includes the AI agent credentials guide and the Cloudflare AI Search agency guide.

Key Takeaways

Strands Decider is Amazon's open ~2B decision model for closed choices with confidence scores.
It swaps a Qwen3.5-2B LM head for a small pointer head and ships under Apache 2.0 with training materials.
Published local latency sits around 115 ms median on an RTX 3090; accuracy claims should not invent a Jev head-to-head win.
Best agency fit: frequent tool, route, and policy checkpoints around a larger generative model.
India and USA teams should treat thresholds, logging, and self-host cost as delivery work, not afterthoughts.

FAQ

What is Strands Decider in plain terms?

Strands Decider 2B is an open-source decision model from AWS Strands Labs. It picks among supplied options and returns confidence scores instead of writing free text. Amazon released it on October 1, 2026, with weights on Hugging Face and training materials on GitHub under Apache 2.0.

How fast is Strands Decider on common hardware?

AWS reports a median of about 115 milliseconds for local decisions on an Nvidia RTX 3090, with small-task medians around 153 milliseconds on an M3 MacBook. Latency rises roughly with task token size. These are measured local figures, not a hosted SLA.

How does Strands Decider compare with TypeSafe Jev?

Jev is a hosted API priced at about $0.042 per million input tokens. Strands Decider is free to download and self-host, with published training data and scripts. AWS has not shown a head-to-head JevBench win over Jev itself, and self-host compute still costs money. The clear Strands edge is openness and local control.

Where should agencies put Strands Decider in an agent stack?

Use it as a cheap checkpoint for bounded choices: tool allow or deny, route to a team, policy yes or no, or premature tool-call checks. Keep generation, coding, and open reasoning on a larger LLM. AWS shows this pattern with the Strands intervention hooks before a tool runs.

Why does Strands Decider matter for India and USA agencies?

Many client projects need frequent agent decisions without paying frontier LLM latency on every hop. India and USA delivery teams can self-host Strands Decider for privacy-sensitive checkpoints, fine-tune on client policies, and document thresholds in the SOW. That pairs well with consulting work on evaluation harnesses and runbooks.

About the author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA software agencies on delivery planning, AI agent pilots, and practical evaluation habits that survive client demos. Debugged Pro helps teams ship reliable software with clear process, not empty hype.