Agents & automation · Part 1 — The AI Dictionary

AI that doesn't just answer — it plans, takes actions with tools, checks results and keeps going until the job is done.

A chatbot responds to one message. An agent works in a loop: understand the goal → plan → act (search, run code, click, edit files, call APIs) → observe what happened → adjust → repeat. Coding agents like Claude Code, Codex and Cursor's agent, research modes that browse dozens of sites, and assistants that operate a browser are all agentic.

Think of it like: The difference between asking a friend for directions and handing them your keys to run the errand.

A model's ability to call external tools — search, calculators, code, APIs — and use the results.

Models can't browse or run code on their own. With tool use, the app tells the model which tools exist; the model replies with a structured request ("call get_weather with city=Austin"), the app runs it, and the result goes back into the conversation.

An open standard for plugging AI apps into tools and data — a 'USB-C port' for AI.

Introduced by Anthropic in late 2024 and since adopted across the industry (OpenAI, Google, Microsoft, Cursor and more), MCP defines one common way for an AI app to connect to an external system — GitHub, Slack, Google Drive, a database, your own API. Build an MCP server once and any MCP-compatible assistant can use it.

Think of it like: USB-C for AI: one standard plug instead of a drawer full of custom cables.

Coordinating multiple AI agents — a lead agent splits work, hands pieces to subagents, and merges the results.

Big tasks overflow one context window. Orchestration breaks them up: a coordinator plans, spins up subagents for research or parallel coding tasks, each with a clean context, then assembles their results. Coding tools like Claude Code and Codex support subagents; research modes use the same idea to read many sources at once.

Repeatable tests that measure whether an AI system does your task well — the unit tests of AI.

Before trusting an AI workflow, you build a set of real examples with known good answers and score the system against them — automatically where possible, sometimes with another model as judge. Change the prompt or model? Re-run the evals and see if you improved or broke something.