RFC: Declarative Skills System for Agent Capabilities

[St3v3a]

I have developed a Python declarative skills system that enables agents to acquire domain expertise through YAML-defined skill files (similar to Claude Skills. This eliminates the need for custom Python code for each data operation, reducing agent implementations from ~350 lines to ~30 lines.

Key Innovations

1. Declarative Item Chains

Skills provide composable data pipelines defined in YAML:

item_chains:
  get_billing_report:
    min_role: AIWORKS_USER
    steps:
      - query:
          tool: execute_template
          template: billing_report
      - transform:
          - parse_mcp: {}
          - extract_rows: {path: "data.tables[0].rows"}
          - field_mapper: {mappings: {model: "Model", cost: "EstimatedCost"}}
      - present:
          output: {by_model: by_model, total: row_sum.cost}

2. Secure Code Execution Sandbox

For complex operations beyond declarative transforms, skills can execute Python in a secure sandbox:

- AST-based validation - Blocks eval, exec, dangerous operations
- Read-only queries - Blocks .set, .drop, .alter in KQL
- Automatic PII filtering - Emails, phones, SSNs, IPs, keys
- Resource limits - Max rows, timespan, timeout controls
- Credential isolation - Secrets never exposed to sandbox

Benefits

- Declarative - Domain experts define operations without Python
- Composable - Pluggable transform operations (parse, extract, aggregate, filter)
- RBAC-integrated - Role validation at skill and operation level
- Testable - Built-in test definitions with mock responses
- MCP-native - First-class Model Context Protocol integration
- Secure execution - Sandboxed Python for advanced operations

Full RFC

📄 https://gist.github.com/St3v3a/e1f593b897bc6f168b24fd8b8712d219

Have been running this in development and would love feedback on whether this pattern would be valuable for the broader Agent Framework community.