Evaluation Runner Template

A builder-readable walkthrough for mapping an AI-agent system from source: identify its architecture, tools, gateways, schedules, state, and review points, then use that map to configure, operate, and troubleshoot the system with a repeatable process.

The Anthropic Agent SDK gives you tool use, streaming, and single-agent loops. Here is an honest map of what it leaves unfinished — and how a solo technical operator closes those

Pydantic AI is useful because it makes typed agent contracts visible. The current evidence supports a control-surface map, not runtime, safety, benchmark, or adoption claims.

Mastra is useful Starkslab evidence because its public repo and docs expose a TypeScript agent framework with clear control surfaces. That supports inspection, not adoption guidance.

Open Computer Use MCP is useful Starkslab evidence because it makes the computer-use runtime boundary visible. That supports inspection, not adoption guidance.

Herdr is useful Starkslab evidence because its repo and docs expose an agent-aware terminal multiplexer: real panes, persistent sessions, state rollups, and CLI/socket controls.