Agentic AI
AWS Agent Toolkit's hidden rules file: the key to making agents actually use available skills
AWS's new Agent Toolkit bundles 300+ APIs and 20+ domain-specific skills, but agents won't proactively load them without a simple configuration file that most users overlook.
The New Stack
AWS released its Agent Toolkit for AWS on May 6, a framework enabling agents to access hundreds of AWS APIs through Model Context Protocol (MCP), paired with packaged domain expertise and architecture guidance. But there's a critical gap between having tools available and having agents use them predictably.
The toolkit's three layers are:
- MCP Server — 300+ AWS APIs, sandboxed Python execution, real-time documentation search
- Skills — 20+ packages of architecture decision tables, service comparisons, and deployment workflows
- Rules file — instructions telling the agent when to load and prefer skills
Most developers, in their rush to test the toolkit, skip that third layer. Without it, agents treat skills like reference books on a shelf—available only if the LLM decides to consult them. With the 17-line rules file, skill loading becomes the agent's first instinct.
The difference it makes#
One developer discovered this after building a serverless backend with the agent's general knowledge alone. The agent produced a solid API Gateway + Lambda + DynamoDB architecture—but never touched the MCP documentation or skills. After adding the rules file, the same agent immediately loaded architecture decision tables before writing code, adding specificity and confidence to its decisions.
The rules file essentially redirects agent behavior from "I can guess this" to "let me check the verified patterns first." For simple applications, the difference is refinement. For complex architectures involving event processing or multi-pattern designs, the skills' flowcharts and service comparisons could reshape the agent's entire approach.
Safety and transparency#
The toolkit addresses security concerns by decoupling agent permissions from developer permissions—you can restrict an agent to read-only even if your IAM role allows deletions. Every API call is logged for auditability, and all skills ship as tested end-to-end workflows.
Setup instructions for Kiro, Claude Code, Codex, and other MCP-compatible agents are in the README. For rules file placement, see AWS's agent toolkit documentation.
