AWS open-sources Strands Decider 2B, a sub-100ms 'decision model' for agent routing

Swami Sivasubramanian, AWS VP of Agentic AI, introduced Strands Decider 2B on October 1 as part of 'a new class of decision models.' Reasoning models generate open-ended text. Decider instead chooses among a fixed set of options, such as which tool to call, which branch to take, or which agent to hand off to, and does it in under 100 milliseconds. At 2B parameters it runs locally, which suits fast experimentation and edge deployment.

In practice, an agent orchestrator frames each routing decision as a constrained classification and lets Decider answer, saving the expensive frontier model for steps that need generation. In multi-step agents, routing calls can dominate both latency and cost, so a cheap, deterministic-feeling selector can cut both sharply. The Apache 2.0 weights and the published training recipe let enterprises fine-tune Decider on their own decision schemas.

Decision models are suddenly a category. AutoTrust released JEV-27B-VL, billed as the first open-weight near-SOTA multimodal decision model. Perplexity open-sourced pplx-decider-27b with a 4-cents-per-million Decisions API the same day. Cloudflare's Clef open-weight decision models drew 428 HN points. VentureBeat quoted a view that AWS 'gives enterprises a credible alternative before Jev builds switching costs.'

Caveats: a 2B model's accuracy on ambiguous decisions is untested against 27B rivals, and AWS published no head-to-head benchmarks in the sources. Watch for Strands Agents SDK integration, Bedrock hosting, and whether 'decider' layers become a standard part of agent stacks or a fad. The other AWS news of the day was a $1B+ Synopsys chip deal.