AI agents
Tool calling and the agent loop, MCP, skills and harnesses, multi-agent architectures, and how to evaluate agents.
Questions I use to pressure-test how LLM agents work as systems — what actually happens when a model “calls a tool”, what the harness around the model is responsible for, how MCP servers and skills plug in, when several agents beat one, and how to measure whether an agent is any good.
Builds on the LLMs & AI systems topic (the model itself); the GCP topic covers Google’s managed agent services, and SDLC & AI tools covers using coding agents day to day.
Subtopics
- Tool use and agent loopsSeniorHow function calling actually works, the agent loop, parallel calls, error handling and runaway-cost guards, and frameworks vs raw API calls.9 concepts
- MCP, skills, harnesses, and guidelinesSeniorThe Model Context Protocol, the skills pattern, what a harness is, and how system guidelines layer on top of the model.9 concepts
- Multi-agent architecturesSeniorWhat a multi-agent system is, the main topologies (orchestrator-worker, pipeline, debate, swarm), how agents are wired in code, shared state and communication, and when not to use one.6 concepts
- Evaluating agents: metrics, ADK, LangSmithSeniorFinal-response vs trajectory vs single-step metrics, tool-call correctness, cost and reliability metrics, and how Google ADK and LangChain / LangGraph (LangSmith, agentevals, openevals) implement them.16 concepts