AWS open-sources Strands Decider 2B to make agent routing decisions on a laptop

Take an agent step whose only job is to choose which of four tools should handle an incoming support ticket. On a general LLM, that step means the model writes a short answer, spending tokens and time to do it ^[2], and the code then parses the answer back into one of the four options. Strands Decider 2B is handed the options and returns one of them, with no other output ^[3]. Each pick arrives with a confidence score meant to show how likely it is to be right ^[4].

AWS pitches the release as a way to speed up agentic AI development more broadly ^[1][16]. What it ships is narrower and easier to test. The head AWS added is about 0.05 percent of the model's parameters ^[17]. The rest is an existing open LLM, tuned with a rank-16 LoRA adapter so the head can score the listed choices directly against answer positions ^[9]. AWS has labeled this release v.20 after repeated rounds of iteration ^[18]. Put plainly, it is a small classifier on top of a small LLM, and a team can check it against labeled examples from its own logs.

So far the evidence is AWS's own, on a benchmark that carries the name of the rival model its team criticises ^[10][11][14]. The coverage does not include the scores, or any test against a large LLM making the same calls.

AWS's own design for using it is a hybrid agent: the decider takes simple, repetitive choices and an LLM takes complex reasoning ^[15]. In that design the confidence score decides the handoff. A pick above a set threshold stands, and anything below it goes to the larger model. That split saves money only if high-confidence picks are right about as often as the score says, so for routing the calibration half of AWS's JevBench claim matters more than the accuracy half ^[10]. If most calls land below the threshold, the agent pays for both models.

Two axes sort the candidate steps. One is whether every option can be written down before the call, since the decider only chooses among choices it is handed ^[3]. The other is what a wrong pick costs. Fixed options with a cheap mistake, such as a tool call that can be retried, are where a trial makes sense now. Fixed options with an expensive mistake, such as a guardrail that lets the wrong request through, can still use the decider behind a high threshold, so doubtful calls go to the LLM. Steps with open-ended output stay on the LLM whatever a mistake costs. For the steps that qualify, the evidence a team needs is its own fallback rate and error rate on real traffic, set against what the step cost when it ran on the LLM.