AWS launched Amazon Bedrock AgentCore Harness in public preview, providing managed infrastructure for AI agents including isolated compute (microVMs), persistent memory, multi-model support, tool connectivity via MCP and AgentCore Gateway, and observability. However, the post argues that infrastructure alone doesn't solve agent governance. It introduces Shape, a ~400-line zero-dependency Python library that adds a governance layer on top: phase lifecycle enforcement (Explore → Decide → Commit), transactional tool calls with automatic rollback, budget-based behavioral gates, proof traces explaining why tool calls were permitted, and a human-readable rule DSL. The two layers are complementary — AgentCore Harness handles runtime, Shape controls what the agent is allowed to do within it.
