Amazon Bedrock prompt optimization adds automated prompt shortening

AWS has added prompt shortening to Amazon Bedrock Advanced Prompt Optimization. The optimizer rewrites prompts to be shorter while checking that output quality holds up. A compression ratio of 0.7, for instance, means a 30% token saving.

The key mechanism is evaluation in the loop. Instead of blindly trimming, the system validates each shortened candidate against the customer's chosen criteria. Those can be LLM-as-a-judge rubrics or custom evaluators implemented as Lambda functions. A shortened prompt is only accepted if it meets or beats the original's scores, which addresses the main risk of compression: quietly degrading behavior on edge cases.

This matters most for agentic workloads, where long system prompts are re-sent on every loop iteration and token costs compound. Community reaction has been positive, with developers reporting meaningful cost cuts in agent loops. It complements other cost levers such as prompt caching, which Google is aggressively discounting with 95% cache pricing on Gemini 4 Argon, and Couchbase's newly added semantic caching for Bedrock.

The caveats are the usual ones for automated prompt rewriting. The quality guarantee is only as strong as the evaluation set, and prompts tuned against a narrow rubric can regress on inputs the rubric doesn't cover. Teams should keep holdout tests and watch production metrics after adopting shortened prompts.