Anthropic and AWS: an Agent Skills suite for creating and managing Automated Reasoning policies in Amazon Bedrock

AWS Machine Learning Blog described how to use an Agent Skills suite to manage the entire lifecycle of Amazon Bedrock Automated Reasoning policies directly from a coding agent. According to the article, Agent Skills are a lightweight open format from Anthropic—a structured package of context that teaches an agent to work with a specific service or domain instead of relying on general training data. Because the format is open, a skill can be installed in any supported agent—the article names Claude Code, Cursor, Codex and Kiro—and activates automatically when the user asks the agent about the relevant topic.

According to the article, an Automated Reasoning check in Amazon Bedrock takes place in two steps: first, a set of foundation models translates the question and answer into formal logic and maps natural language to variables in the policy; then an SMT solver (Satisfiability Modulo Theories) checks this logic against the rules and returns a verdict. This verification step is mathematically sound—if the translation is faithful, the verdict is correct—and the result is explainable because each verdict lists the specific rules that support or contradict it. Rules are written in a subset of SMT-LIB.

The suite described contains six Agent Skills, one for each phase of the policy lifecycle: builder creates a policy from a source document and extracts rules and variables, reviewer checks quality and fidelity reports and identifies conflicting rules, unused variables and bare assertions, tester generates scenarios and runs question-answer tests, debugger diagnoses errors and fixes the policy, deployer creates a numbered policy version and attaches it to a guardrail, and validator checks responses through the ApplyGuardrail API, including a loop that rewrites an incorrect response based on the violated rules. Each skill has a standard structure—a SKILL.md file with instructions, a references/ folder with details and a scripts/ folder with executable Python scripts that call the relevant API and support the --help and --dry-run flags.

The article documents a walkthrough of the entire process using a short HR policy on parental leave eligibility. According to the article, use requires an AWS account with access to Amazon Bedrock in a region where Automated Reasoning checks are available, the appropriate permissions and Python with uv to run the scripts. Details can be found in the source article.