By the time I wake up most mornings, my AI system has already planned my day and worked through my inbox.
When I show people this, they often ask the same question: "Which model do you use?" It is the least important part of the answer.
I have spent the better part of a year building what I would call an agentic operating system. It has six parts, and the easiest way to explain them is as a body. This article walks through those six parts, why the model matters less than people think, and what the same anatomy means for a team that wants AI agents to do real work.
The Six Parts of an AI Agent
An AI agent is a system that can take a task, work out the steps, use your tools and data to carry them out, and report back. The language model is only one piece of that. Here is the whole body.
The brain: the language model
The model reads, reasons and writes. It gets all the headlines. It is also the only part I rent, and I swap it whenever a better one ships.
The heart: context
Everything the system knows about my clients, my projects, my preferences and how I work. This is the biggest piece by far. Without it, the smartest model is a brilliant intern on day one: capable, but useless to you specifically.
The hands: connectors
Email, calendar, CRM, databases. Connectors let the agent read from and act in the systems where the work actually happens. No hands, no work.
The tool belt: skills
Procedures the system has learned once and keeps: how we write invoices, how we prepare meetings, how we review outbound text. A skill turns a good result into a repeatable one.
The legs: automations
Routines that run on a schedule, whether I am at the desk or not. This is what makes the morning briefing appear before I am awake.
Wrapped around everything: security
The rules about what the system may never do without asking me first. It is not a sixth organ so much as the skin around the other five.
The model sits in the head, but most of the body is context, connectors, skills and automations. Security wraps around all of it.
Why the Model Is the Least Important Part
Here is the part most people get backwards. Five of these six parts compound over time. The brain is the one piece that does not, because it is interchangeable.
Every week the context gets richer, because the system records what happened and what was decided. Every new connector opens another system to work in. Every skill I write once is used on every future run. Every automation keeps running while I do something else. And the security rules get sharper each time I see something I do not want to happen again.
The model is different. When a better one ships, I switch, and everything I built around it keeps working and usually gets better on the spot. Nothing I invested in the old model carries over, and nothing needs to.
When a company tells me they are waiting for a better model before they start, I hear: we are waiting for a bigger brain, for a body we have not built.
The model is rented and swapped. The other five parts stay, and they keep growing.
What the Six Parts Mean for an Expert Team
The same anatomy applies when an agent works for a team instead of one person. The words change a little. The order of importance does not.
Brain: whichever capable model your company is allowed to use. Choose it with your data rules in mind, then treat it as replaceable.
Heart: the knowledge your experts carry around. Past cases, decisions and the reasons behind them, templates, the exceptions everyone knows about but nobody wrote down.
Hands: access to the systems the workflow touches, with the same permissions a colleague doing that work would have. Not more.
Tool belt: your team's way of doing things, written down as procedures the agent follows. How you check a document, how you structure a report, what a good answer looks like.
Legs: the recurring moments where work should start by itself. A weekly report, an incoming request, a deadline that is coming up.
Security: which data may go where, which actions need a person's approval, and who reviews what the agent produces.
Notice where the value sits. The model is the same one your competitors can rent. The context and the procedures are yours, and they come from the people who already know the work.
Where to Start: Context and Connectors on One Workflow
You do not build all six parts at once, and you do not build them for the whole company. Start with one workflow your team repeats, and build the parts that workflow needs.
In practice that means context and connectors first. Collect what an experienced person would need to know to do the task well, and give the agent access to the one or two systems where the task lives. Then write down the procedure as a skill. Automations come last, once the team trusts the result enough to let it start without being asked.
This order also keeps the work honest. If the agent produces poor results, the cause is almost always missing context or a missing connection, not a weak model. That is useful to know before anyone spends money on a bigger one.
This is how we approach agent projects with clients: one workflow, built with the people who know it. You can read more on how we work .
What Stays With People
My own system can plan my day and draft replies to my email. It is not allowed to send an email on its own. That line is written into the security layer, and it stays there however good the drafts become.
Every team needs its own version of that line. Decide which actions the agent may take alone, which ones it prepares for a person to approve, and which ones it never touches. Those decisions belong to the people who are accountable for the work, not to the model and not to whoever built the agent.
Drawing those lines early is not a brake. It is what makes it safe to give the agent more hands later.
Build the Body, Rent the Brain
The model will keep getting better without any effort from you. The context, the connections, the procedures, the routines and the rules will not build themselves. They are also the parts that make an agent useful to your company specifically.
So the useful question is not which model to wait for. It is: which of the six are you actually building?
Frequently Asked Questions
- What are the main components of an AI agent?
- A working AI agent has six parts: a language model that reasons, context about your business, connectors to your systems, skills that encode your procedures, automations that start work on a schedule or trigger, and security rules that define what the agent may do without a person's approval.
- How do AI agents work?
- The agent receives a task, uses the language model to plan the steps, pulls in the context it needs, uses connectors to read or act in your systems, and follows stored procedures for how the work should be done. Security rules decide which steps it can take alone and which need approval.
- Should we wait for better AI models before building agents?
- No. The model is the one part you can swap at any time. Context, connectors, skills, automations and security take time to build and keep improving, and a better model makes all of them more valuable the day it arrives.
- Which part of an AI agent should a company build first?
- Start with context and connectors for one recurring workflow: what an experienced person needs to know, and access to the systems where the work happens. Add skills next, and automations once the team trusts the output. Decide the security rules before the agent gets access.
How we work · Work · Insights