NetApp
/
Generate Practitioner
/
How an Agent Actually Works
NetApp Iterate.ai
NetApp Sellers & Partners
Joint Educational Series
Field Brief · The Building Blocks

How an Agent Actually Works

Every brief in this series says “agent” and “tool” as if you already know what they mean. This is the one that explains them.

What you'll learn
What actually happens between your question and its answer
What a “tool” is, and why a model needs one at all
The one standard that lets an agent use almost any tool

The core argument

A chatbot has a conversation. You ask, it answers, from what it already knows. An agent has a job list. It can stop mid-answer, go get something it does not know, and come back — then decide if it needs to go get something else before it is done.

Every one of those trips is a tool call. A model can write and reason, but it cannot check today's date, look up a record, or send an email by itself. A tool is what gives it hands — one specific, narrow action it can trigger.

A tool can be almost anything a business already uses. If a person could look something up or take an action there, an agent can be handed a tool that does the same thing — including searching the open web.

Email (e.g. Gmail)CalendarSpreadsheet (e.g. Excel) Accounting (e.g. QuickBooks)A database (e.g. Snowflake)The open web Messaging (e.g. Slack)CRMSupport ticketsA calculator
One question, from prompt to answer
1
You ask
That is the prompt
2
Agent decides
Do I need a tool for this?
3
Tool runs
Looks something up, or does something
4
Answer
Comes back, maybe after several trips

Steps 2 and 3 repeat as many times as the job needs — the agent decides, one trip at a time.

A worked example

“What is my copay for tomorrow's visit?” The agent makes two tool calls to find out.

Look up insurance tier
Look up appointment type
“$40”

Key facts

A prompt is the instruction. A tool is the action. An agent decides when to use which.
Most of what an agent does is invisible to the person who asked the question.
A tool can be as simple as a calculator or as complex as a full database query.
Every tool call is a place data can leave — or stay inside — your walls.
NetApp | Iterate.aiHow an Agent Actually Works · 01 / 02

The standard that lets it plug in anywhere

Without a shared standard, every agent needs custom wiring for every tool it might use — one connector per pair, rebuilt each time either side changes.

MCPModel Context Protocol — fixes that. It is a shared job description any tool can implement. Once a tool speaks it, any agent that also speaks it can find that tool, see what it does, and use it. Think of it like a wall socket instead of a bare wire: once something has the right plug, it works in any socket built to the same standard.

MCP is not one company's product — it is an open specification. MongoDB, for one, now ships its database as an MCP tool, so any MCP-speaking agent can query it with no custom connector.

41%
Of software organizations already run MCP-connected tools in production (Stacklok, 2026)
10,000+
Active public MCP tools an agent could potentially reach (Anthropic ecosystem update)
970×
Growth in monthly adoption of the standard in its first eighteen months
Why this is a governance question too

Every tool a model can call is also a door. A standard that makes doors easy to add makes it easy to add too many. An unreviewed tool is the same risk as an unapproved app — just faster, and running all day.

Key facts

MCP decides how a tool is described and called — not what it is allowed to do. Permissions are a separate decision.
A tool speaking MCP can be found and used by any agent that also speaks it, with no custom code.
The same rule applies here as to every model call: know what it is, and log what it did.
The standard isn't the safeguard

MCP is the plug. It is not the policy. Standardizing how an agent finds and calls tools makes adoption fast — which is exactly why deciding, deliberately, which tools it is allowed to use matters more, not less.

Further reading

Full paper, 10 pages — iterate.ai/partners/netapp/papers. On deciding which model handles a job, see What Is a Harness?; on seeing every call after the fact, You Cannot Govern What You Cannot See.

About this series

A joint educational series on private AI and the AIPod Mini. NetApp® — the governed data-control layer. Iterate.ai — the private intelligence layer.

NetApp AIPod Mini
NetAppIterate.ai
v1.1 · Aug 11, 2026
NetApp | Iterate.aiHow an Agent Actually Works · 02 / 02