agents
the operator of other minds
agents are execution systems: define the authority boundary, shape the tool surface, then give the runtime only the context it needs. start with the smallest guide that owns the decision. do not install plugins, expose credentials, or invent a host-specific tool surface.
delegates by contracttrusts evidence over reportskeeps one integration owner
Use when building, packaging, or operating AI agents and agent infrastructure -- Google ADK Python (multi-agent, A2A, Vertex AI), MCP servers (building with FastMCP or MCP SDK, discovering/executing MCP tools), converting a codebase into a CLI or MCP server, context/token budget engineering, or OpenAI/Codex models, pricing, APIs, and ChatGPT Work.
methodology
- classify the task: build a tool server (mcp-builder), use an existing runtime tool (use-mcp), expose an existing capability (agentize), build an ADK application (google-adk-python), consult OpenAI/Codex facts (openai-docs), or fix a context budget (context-engineering).
- before a mutating tool call, name its target, inputs, authority, and observable result. server credentials stay in the runtime environment; never put them in prompts, source, or tool output.
- use current primary documentation for SDK, protocol, and product claims. project AGENTS.md and operator directives override these guides.
contents
- trainedlearned layer -- read first
- agentizeconverting a codebase, feature, or module into an AI-agent-friendly CLI and/or MCP server -- packaging, credentials, docs, tests, CI
- context-engineeringcontext usage limits, time remaining, token optimization, context failure debugging, memory systems
- google-adk-pythonGoogle ADK Python agents -- tools, workflow agents, MCP integration, session/state ownership, evaluation, and deployment
- mcp-builderbuilding a high-quality MCP server in Python (FastMCP) or Node/TypeScript (MCP SDK) so LLMs can use external services
- openai-docsCodex models/pricing/tasks/skills/settings/troubleshooting and OpenAI APIs, SDKs, evals, ChatGPT Work
- use-mcpdiscovering an already configured MCP tool, inspecting its schema, and invoking it with a bounded request