What is an AI agent?

Most pages selling an agent present it as one indivisible product: you buy it or you do not. Convenient for selling, misleading for deciding — an agent is not a block. It is three separable things.

The harness is the program that runs the loop: system prompt, context management, guardrails, tool routing, error recovery.

The model is the LLM. The layer everyone talks about, and the only one that genuinely swaps.

Tools and skills are what the agent can do and what it knows how to do. Two different things — more on that below.

Take the conclusion before the detail: the most-discussed layer is the most replaceable, and the other two are the ones you own.

An agent demo fits in a prompt. An agent in production fits in the harness — and that is precisely what demos leave out. If you take one term away from this page, take this one.

• The system prompt — the role, the limits, the tone, what the agent may not do.
• The loop — when to call a tool, when to stop, how many turns before giving up.
• Context management — a long run overflows the model's window. Something has to summarise, prune and decide what survives. Done naively, the agent forgets the essentials and keeps the small talk.
• Guardrails — what is checked before an action leaves, and what requires a human.
• Error recovery — an API that falls over, a tool returning something unexpected, a model inventing an argument. In production this is Tuesday.

This is software, not prompting. It is also reusable: one harness serves several agents. And it is why "just get a subscription" is an incomplete answer — the subscription gives you the model, not the loop.

An LLM answers. An agent uses an LLM as one component among several. That is the whole distinction, and it is why "which model is best" is rarely the decisive question.

Claude, GPT, Gemini, an open-weight model behind an agent such as Hermes. They are not equivalent at everything, but they play the same architectural role: they reason and decide which tool to call.

Swapping models is the cheapest change available — not a free one. Prompts need retuning, tool-calling reliability differs between models, context windows differ. Count days, not months. Compare that with rebuilding tools, which is counted in weeks.

No general assistant ships telephony. Turning an agent into a phone line means adding a carrier, speech-to-text, text-to-speech, and a latency budget that decides whether the result is usable or painful. That is not a model weakness: it is an interface layer, and it needs infrastructure.

A tool is a function the model can call: read a calendar, raise an invoice, query a database. Deterministic — the agent does something.

A skill is a written procedure loaded when it is needed: how a quote is drafted at your company, in what order, with which clauses. Not executable — the agent knows how.

Conflating them is expensive, because you end up coding what should have been written, or writing what should have been executed.

MCP, and why it changes the calculation

The Model Context Protocol is the standard way to attach a tool to an agent. The major agents have converged on it, and the consequence is simple: a tool written once works with the next model and the next harness.

Move from Claude to ChatGPT, or to an open-weight model you host yourself: your tools come with you. Only the harness is refitted, and it is built to be.

This is what makes coherent a position that might look contradictory — saying "do not depend on one vendor" while working with commercial models. We do not sell a model. We build tools that outlive the model.

Three layers describe an agent. It takes three more for one to stand up in production.

Memory and knowledge. Short term is the context window. Working memory is files. Long term is a store or an index. This is not harness: it persists between sessions and it is your asset. Whoever holds the memory holds the value.

Interface. How people and systems reach the agent: a chat, Slack, email, the telephone, a scheduled job, an API. The same agent behind two interfaces is two different products to the person using it — and this is where the voice question actually sits.

Governance. Whose credentials the agent acts with, what it is permitted to do, what is logged, how you know it still works. This separates a demo from a system, and it is the question to ask before signing: who holds the access?

There is no general answer. There is a test.

Buying wins when the need is standard, when it has to run this week, and above all when the product carries infrastructure you do not have — telephony first among them. A packaged product has already solved problems you would rather not solve.

Building wins when the process is yours rather than the industry's, when the data cannot leave, and when the horizon is beyond two years. At three years, a per-seat subscription often costs more than paying once would have.

The test hides a third option, which is the common one in practice: buy the substrate, build what distinguishes you. Take a model on subscription and invest in the tools and the harness — the two layers that stay yours when the rest changes.

Compare totals, not headline prices.

Packaged: the per-seat monthly subscription, initial configuration, connectors billed separately, and the exit cost — how long to get your data and your processes back on the day you leave.

Built: design, tool development, the model subscription you still owe, hosting, maintenance. Heavier up front and flatter afterwards.

The figure almost nobody calculates is the exit cost. It is invisible in year one and decides everything in year three.

Without us, and without a budget.

1. Pick a repetitive, boring task somebody does by hand several times a week. Not the most strategic one: the most regular one.
2. Write the procedure down as though explaining it to a new hire. If you cannot, no agent will manage either — and you have just learned something useful.
3. Do it by hand with an assistant, copy-pasting, for a week. You will find the friction before writing a line of code.
4. List the tools that would need connecting to remove the copy-pasting. That list is your specification.

Four steps, a few hours, and whole projects avoided.