Amazon releases a small open model for agent decisions

Strands Agents introduced Strands Decider 2B on October 1. The project describes it as a small open-source decision model for fast experimentation and local development. Unlike a general language model, it is meant to select from defined choices, such as yes-or-no checks, and can sit beside an agent that still uses a larger model for other work.

For makers, this is a useful split in the agent stack. A narrow decision step can be easier to test, faster to run, and less likely to produce an open-ended action than asking one model to plan and act at once. It does not remove the need for permissions or a human review path.

Try it on one bounded gate: tool eligibility, input validation, or escalation. Compare its decisions with a simple rule and your current model, then record false approvals, false blocks, latency, and local operating cost before widening its role.