Amazon Bedrock has released Advanced Prompt Optimization, a managed service that automates prompt engineering across multiple models. For teams wrestling with prompt sensitivity or model migrations, this tool replaces manual iteration with a metric-driven feedback loop.

At its core, the optimizer accepts a prompt template with variable placeholders, example inputs, ground truth responses, and an evaluation metric. It then runs a closed loop: sends prompts to up to five models simultaneously, scores responses, and rewrites prompts to maximize the chosen metric. The output includes original and optimized templates, evaluation scores, cost estimates, and latency figures.

Three evaluation methods are available. For deterministic tasks like JSON matching or F1 scores, you can deploy a Lambda function with custom Python scoring logic. For open-ended tasks such as summarization, an LLM-as-a-judge pattern uses a configurable rubric and a judge model (default is Claude Sonnet 4.6). A lighter option is steering criteria—free-form natural language constraints like brand voice or format rules—which the system incorporates into a default judge prompt.

Multimodal inputs are supported: prompt templates can reference PNG, JPG, or PDF files from S3, enabling optimization for document and image analysis workflows.

This is particularly useful for migration scenarios. Select your current model as a baseline, add up to four target models, and the optimizer will test for regressions on known cases while improving underperforming tasks. The entire job runs asynchronously, with results written to an S3 output location.

Pricing follows standard Bedrock inference token rates. Available in US East, US West, Canada, Europe, Asia Pacific, and South America regions. The API is CreateAdvancedPromptOptimizationJob.

Source: AWS