How AI Agents Actually Work: APIs, MCP, and the New Deal Stack

Most of the debate about AI agents skips the part that actually matters. People argue about whether agents will replace jobs, or whether software is dead, while quietly avoiding a simpler question: how does an agent do any of this?

Not philosophically. Mechanically. How does an agent pull data from one system, process it, and push the result somewhere else? How does it reach your email, your data providers, your deal platform?

There is a new vocabulary behind the answer. API. MCP. CLI. You hear these terms in every serious conversation about enterprise AI now, and if they do not mean much to you yet, they will soon. Here is what they are, in plain language, and why they decide who gets value from AI and who watches it route around them.

What is an AI agent?

An AI agent is software that uses a language model to carry out multi-step tasks on its own. A chatbot answers a question. An agent takes actions: it reads a document, extracts the figures, compares them against a database, and drafts the output, chaining steps together toward a goal.

That autonomy is only as valuable as the agent’s reach. An agent that cannot touch your systems is a very articulate thinker with no hands. Everything below is about the hands.

What is an API?

API stands for Application Programming Interface. In plain English, it is the way one piece of software talks to another.

When you log into a data provider and pull up a company profile, you use their interface. You click buttons, scroll through tables, read charts. That interface was designed for you, a human. An API is the equivalent designed for software. Instead of clicking a button to get a company’s revenue, an agent sends a structured request and gets back structured data. No screen. No clicks. A direct connection.

Every tool you use has an API, or should. Your email. Your CRM. Your data providers. Your deal platform. When people call a product “API-first,” they mean it was built so other software, including agents, can work with it natively.

This is also why seat-based software pricing is under pressure. If a product was built for humans to click through and has no API for agents to call, it is invisible to the new workflow. The agent cannot use it, so it routes around it. The products that last will be the ones that work for both humans and agents.

What is the Model Context Protocol (MCP)?

Here is the catch. Every API speaks a slightly different language. One data provider’s API works one way, another’s works another way, your CRM works a third way. Historically, connecting an agent to ten tools meant ten separate custom integrations, each one fragile and each one a maintenance burden.

MCP, the Model Context Protocol, is an open standard that removes that problem. It gives AI models one universal way to connect to external tools and data sources. One protocol. Any tool that supports MCP can plug into any agent that supports MCP, from any provider.

The clearest sign of where this is going: Anthropic shipped MCP connectors to FactSet, MSCI, and S&P Global. Agents can now pull institutional-grade financial data through a standardised connection, and the same protocol that reaches your email reaches your market data. Because the standard is open, no single vendor owns it.

Why “USB for AI” is the right analogy

Before USB, every device had its own connector. Printers, keyboards, cameras, all different cables. USB standardised the connection, and suddenly any device worked with any port.

MCP is doing that for AI. One standard connection layer instead of a tangle of bespoke integrations. That is why it matters far beyond any single announcement: standards are what let an ecosystem compound, because every new tool that adopts the standard instantly works with everything else that already has.

What this looks like in practice

Picture one window. From an AI assistant you pull recent emails tied to a live deal. You send that context to a deal platform that already holds the CIM and supporting documents in structured form. Its agent extracts the audited financials, benchmarks them against comparable deals in your database, and pushes the structured output back. You combine everything into a draft investment committee memo. One session, no switching tabs.

Every step works because each tool exposes an API and MCP provides the standard connection between them. The email, the deal platform, the market data, all reaching the same agent through the same protocol. This is not a distant scenario. The pieces exist today.

Do AI agents replace software, or sit on top of it?

Mostly they sit on top of it. An agent gets work done by calling software through its API, so the value shifts toward the tools that expose clean, agent-friendly connections. The layer at risk is software built only for human eyes: a dashboard with no API, sold per seat, that an agent cannot use and therefore ignores.

That reframes the “is software dead” question. Software is not dead. Software that cannot be reached by an agent is the part that struggles.

The real bottleneck is governance, not models

It would be easy to pretend this is frictionless. It is not. Plenty of teams want to connect their AI tools to their document stores, data providers, and internal systems, and cannot yet, because compliance has not approved it or IT has not scoped it. Those access controls are legitimate.

The larger bottleneck over the next year is governance, and specifically lineage. When an agent pulls a gross margin figure into a memo, where did it come from? Which document, which version? Was it the audited financials or the management case? If you cannot trace an answer back to its source, you cannot trust it, and if you cannot trust it, you cannot put it in front of an investment committee.

The answer is not to control the model. It is to control the data layer, so every piece of information flowing through an agent workflow carries a clear source, timestamp, and chain of custody. Get that right and agents become dependable colleagues. Skip it and you get fast answers you cannot defend.

Where this is heading

Connectivity compounds. Every new tool you connect and every new data source you plug in makes the agent more capable. A firm that wires up five systems this year runs circles around a firm still copy-pasting between tabs next year.

So the vocabulary is worth learning. API, MCP, CLI. Understand the connections, because the firms that wire up first will not just be more efficient. They will be doing a different kind of work, on the same deals, at the same desks, in a fraction of the time.