No jargon for its own sake. If you run a small business, this is the background you need before you hand any task to an agent.
A chatbot answers a question and stops. An agent is given a goal and works toward it: it plans steps, uses tools (search, documents, apps), checks its own work, and reports back. The practical difference is who does the next step. With a chatbot, you do. With an agent, it tries — which is exactly why a human still needs to check anything that leaves your business: an email sent, money moved, or a promise made to a customer.
Think of it in three parts: the model that reasons, the tools it is allowed to use, and the limits you set. The model gets the attention, but the tools and limits decide whether the result is useful or dangerous.
MCP (Model Context Protocol) is an open standard for connecting an AI assistant to tools and data — your documents, your calendar, your business apps — through one shared way of asking, instead of a custom integration for every app. For a business owner, the point is practical: a tool that speaks MCP can plug into more of your workflow, and you can swap one tool for another without rebuilding everything. MCP is a protocol, not a product, and it does not make an agent trustworthy by itself — what it can touch is still your decision.
A2A (Agent2Agent) is a protocol for agents to work with each other: one agent discovers what another can do, hands it a task, and gets a result back. Imagine a research agent handing a draft to a scheduling agent without you copying text between them. It matters because real business work is rarely one agent's job — it is a chain of them. The same rule applies as with MCP: every handoff is a place where a mistake can travel, so the chain needs a human checkpoint before anything external happens.
When a task passes this checklist and you are connecting a real agent to real accounts, work through Before You Give an AI Agent Access first — the permission ladder and the minimum safe setup.
Run through this before handing a task to an agent. If you cannot answer one, fix that first — it is cheaper than cleaning up after a bad run.
If the steps change every time, an agent will guess. Start with tasks you could write down as a checklist.
An agent can only use what it is connected to. If the information lives in someone's head or a locked system, connect it or pick another task.
A wrong internal draft is cheap. A wrong invoice, email, or booking is not. The higher the cost, the earlier the human check must come.
Name the checkpoint before you start: who reviews, what they look at, and what stops the workflow if they say no.
Pick one measure — time saved, errors caught, tasks finished — and compare it against doing the task by hand.
One task, one agent, one week. Prove it, then expand. Five new automations at once is how small businesses end up with five new problems.
This site is itself an example. bizn.org was built, and is now run day to day, by a human working with a Muse AI agent — for $0 in running costs, with the human directing and the agent building and checking its work live. Read how this site was built →
Ready to pick a tool? Browse the directory → · Want to see agents in action? Can an agent do this? → · Building an agent? Capability cards and the validator →