Simulation

Note

Simulation is in public preview. Features and APIs may change before general availability.

Simulation lets you evaluate your agent against dynamic, multi-turn conversations generated by LLM-backed actors. Instead of replaying fixed scripts, you define actor profiles with personas and goals, and the simulation engine drives realistic conversations with your agent.

Simulated scenarios integrate with the same dataset runners used for predefined scenarios. You include simulated scenarios in your dataset, set a simulation configuration on the runner, and the runner handles the actor lifecycle internally.

Simulation types

Amazon Bedrock AgentCore supports the following simulation types:

• User simulation: An LLM-backed actor plays the role of an end user interacting with your agent. The actor has a persona (traits, context) and a goal, and generates realistic user messages until the goal is met or a turn limit is reached. Use user simulation to test how your agent handles diverse user behavior, open-ended conversations, and edge cases at scale.

Prerequisites

• The simulation extra: pip install 'bedrock-agentcore[simulation]'

  This installs the actor simulation engine used to generate dynamic conversation turns.

• AWS credentials with permissions to invoke the actor model via Amazon Bedrock. The actor runs as a separate LLM call, so the credentials used by the runner must have bedrock:InvokeModel permission for the model specified in SimulationConfig.model_id.

• All dataset evaluation prerequisites apply.

Topics

• User simulation