// plain-language glossary
Nine operating terms for AI agents. Each entry pairs a plain analogy with a diagram that shows the moving parts.
01
Like a teammate with a job, tools, and guardrails.
An agent is AI that works through a task in steps. It decides what to do next, uses approved tools, checks what happened, and stops when it can answer or needs human input.
The difference from chat is action: an agent can keep state and work through a loop.
02
Like the approved apps and instruments a worker can use.
Tools are actions an agent is allowed to call: search, files, browser, code, database, calendar, or an internal API.
Tools create real effects, so they need permissions, limits, and logs.
03
Like a recipe card for repeatable work.
Skills are packaged instructions an agent loads for a job. They can include steps, examples, scripts, templates, and done checks.
They turn good practice into a reusable method, not a one-off prompt.
04
Like USB-C: one bus can connect to many different devices.
MCP gives an AI app one standard connector for outside systems: files, databases, browsers, business tools, and services.
It is the bus, not the safety policy. Access still needs rules and approval.
05
Like one accountable operator handling the whole job.
One agent plans, chooses tools, executes steps, checks results, and writes the answer. It is both planner and worker.
Strong for focused work. Weaker when the job needs separate experts or review.
06
Like a delivery team with specialists and a lead.
Multiple agents share the work. One may plan, another researches, another builds, another reviews, and another reports.
The hard part is coordination: ownership, handoffs, decisions, and quality gates.
07
Like a review panel where each member has a defined lens.
Several agents inspect the same evidence from different angles: risk, architecture, cost, delivery, or quality.
A council should produce a decision, dissent if needed, and a next action.
08
Like a governed team notebook and archive.
Memory stores useful context from past work: decisions, preferences, facts, artifacts, and lessons.
It needs rules for what gets saved, corrected, retrieved, and forgotten.
09
Like answering with an open browser and a marked-up reference pack.
RAG makes the AI look up fresh or private information first, then use those sources to answer.
Good RAG separates search from answer writing and shows which sources shaped the result.