#

Key takeaways

• AWS published a technical guide on connecting decision models to Amazon Bedrock AgentCore Evaluations.

• AgentCore Evaluations supports on-demand, batch, and online evaluation.

• The guide uses code-based evaluators that run in AWS Lambda.

• It shows two backends: Jev and Strands Decider 2B.

• AWS says users should validate each question on their own traffic before using a check as a release gate.

What AWS published

AWS published a technical guide on November 4, 2026. The guide explains how decision models can connect to Amazon Bedrock AgentCore Evaluations through code-based evaluators. It presents this as a technical pattern for evaluation workflows, not as a product launch or a customer case study.

The post focuses on how evaluator logic can be extended with code. It says code-based checks run in AWS Lambda and can call external systems, models, or APIs. That makes them different from built-in evaluators, which are already part of AgentCore Evaluations.

What AgentCore Evaluations supports

AgentCore Evaluations supports on-demand, batch, and online evaluation. AWS says it includes 15 built-in evaluators. It also supports custom language-model judges and code-based checks.

The guide places decision models alongside broader language-model evaluations. AWS says language-model explanations remain useful when judgments need detailed reasoning. Decision models are presented as a narrower option for questions with fixed answer sets.

The two decision model backends

The tutorial integrates two backends. One is Jev, a hosted model from TypeSafe. The other is Strands Decider 2B, an open-source decision model with two billion parameters that users host themselves.

AWS says Jev requires a TypeSafe API key. It also says Strands Decider needs a reachable server. Those are operational requirements for the tutorial setup. They are part of the implementation details shown in the guide.

How decision models work in the guide

Decision models choose from a fixed set of answers. They return probabilities rather than free-form explanations. AWS says the supported question types include yes or no, choice among up to 255 options, and ordered scores with two to ten levels.

The guide says these outputs can support thresholds and screening for narrow evaluation questions. That makes them useful when the task is to classify or score a response, rather than explain it in detail. AWS also shows how to convert evaluator inputs into decision-model requests and then translate returned probabilities into AgentCore results.

How AWS frames the role of these models

AWS positions decision models as additions to broader language-model evaluations. The guide does not replace language-model judges. Instead, it shows a way to combine structured decision outputs with the reasoning style of language models when needed.

This framing matters because the two approaches serve different evaluation needs. Decision models are suited to fixed-answer checks. Language models remain useful when the evaluation requires more detailed reasoning or explanation.

Example workflow in the tutorial

The post includes a sample HR assistant and labeled scenarios. AWS presents them as a demonstration of the workflow. It does not describe them as production performance or as a new customer deployment.

The tutorial emphasizes validation before operational use. AWS says users should validate each question on their own traffic before using a check as a release gate. That is the clearest operational caution in the guide. It suggests that the same evaluator may behave differently depending on the user's own inputs and thresholds.

Practical implications for evaluators

The guide shows a structured path for teams that want to use decision models in evaluation pipelines. It combines code-based logic, model outputs, and AgentCore evaluation results in one flow. The result is a setup that can support narrow checks with probabilities and thresholds.

The post is careful about scope. It describes a technical integration and a sample workflow. It does not claim production results, and it does not announce a broader deployment. It also makes no Moroccan availability or deployment announcement.

For readers, the main value is the implementation pattern. AWS is showing how to connect fixed-answer decision models to an evaluation system that already supports multiple modes and evaluator types. The guide gives a concrete example of how structured model outputs can fit into a larger evaluation process.