이미지: AWS ML Blog
Summary
- AWS has released six open-source Agent Skills that let coding agents run the entire lifecycle of Amazon Bedrock Automated Reasoning policies
- Agent Skills is a lightweight open format proposed by Anthropic that can be installed across various coding agents such as Kiro, Claude Code, Cursor, and Codex
- Automated Reasoning checks determine whether AI responses comply with policy using SMT solver-based formal logic verification rather than statistical sampling
- 발행처
- AWS ML Blog
- 발행일
- 2026-08-06
- 작성자
- Adewale Akinfaderin
- 스킬 구성
- 정책 라이프사이클 단계별 6종
- 호환 에이전트
- Kiro, Claude Code, Cursor, Codex
- 정책 작성 언어
- SMT-LIB 부분집합
- 검증 엔진
- SMT 솔버(Satisfiability Modulo Theories)
Moving Console Work into Code
AWS has released a set of open-source Agent Skills that let coding agents run the Automated Reasoning policy lifecycle in Amazon Bedrock end to end. The announcement was published on the AWS ML Blog under the byline of Adewale Akinfaderin. Previously, creating and validating policies was specialized work done in the Bedrock console, but this new skill set turns it into an engineering workflow that coding agents can perform repeatedly.
What Automated Reasoning Checks Verify
Automated Reasoning checks work in two stages. First, a foundation model (FM) translates a question and answer into formal logic, mapping them to the policy's variables. Then an SMT (Satisfiability Modulo Theories) solver checks this logic against the policy rules and issues a verdict. Because this is formal logic-based verification rather than statistical sampling, the mathematical soundness of the verdict is guaranteed as long as the translation is accurate. Policy rules are written in a subset of SMT-LIB, the standard input format for automated theorem provers.
Policy Lifecycle Split into Six Skills
The released skill set consists of six skills, each mapped to a stage of the policy lifecycle. Each skill consists of a short instruction file and an execution script that calls the Bedrock Automated Reasoning API, and they share a common reference document describing the API structure, detection types, and rule syntax to provide consistent guidance.
| Stage | Content |
|---|---|
| Rule extraction | Extract rules from source documents |
| Review | Review service-generated rules |
| Test writing | Write tests reflecting real user queries |
| Debugging | Diagnose failure cases |
| Deployment | Deploy versioned policies behind guardrails |
| Validation | Verify behavior of deployed policies |
Why Agents Are Suited to This Work
Agent Skills is a lightweight open format proposed by Anthropic — a structured context package that extends coding agents with expert knowledge and workflows for a specific service or domain. Rather than relying on incomplete or outdated general training data, agents follow the verified patterns, common pitfalls, and step-by-step procedures embedded in a skill to generate correct API calls. Because the format is open, it can be installed on any agent that supports it, and AWS cited Kiro, Claude Code, Cursor, and Codex as compatible examples.
AWS explained that policy authoring involves a lot of detailed, repetitive work — from extracting rules to handling API constraints — and that such tasks, with their clear rules and failure modes, are well suited to coding agents given proper guidance. It added that teams adopting Automated Reasoning checks want to manage the lifecycle through code precisely because it offers repeatability and ease of review.
metallab.ai continues to cover cases where coding agents are combined with cloud services.
Remaining Questions
AWS said it shared insights gained about the service's behavior while applying this skill set to actual Bedrock deployments, but the excerpted material did not include specific execution results or performance metrics, so further confirmation is needed.
