The most useful mental model for an AI agent was not invented by an AI researcher. It was invented by every manager who ever onboarded a new hire.
A new employee shows up on day one. They have a body — a desk, a laptop, a login. They have a brain — what they know, what they were trained on, how they think. They have hands — the tools they were given access to. Over time they accumulate memory of what works at this company and what doesn't. And eventually, if you're lucky, they develop something harder to define: a soul — the personality, the style, the way they write an email to an upset customer that fits your brand.
Every one of those layers is a design choice for an AI agent. Skip a layer, and you get a chatbot. Build all five, and you get an employee.
The body is the UI. The chat window. The voice on the phone. The form the customer fills in. The dashboard the operator uses. The Slack channel the team monitors. The Telegram thread.
Pick a body you can change. Most teams start with a chat surface — a sidebar with streaming responses, dynamic tool panels, context windows that scroll. Some add voice. Some add SMS. The body should be swappable without re-engineering the brain. The body is not the product. The body is the front door.
A useful test: can your team ship a new body in a week without touching the brain? If yes, the body is properly decoupled.
The brain is the model. In 2026 the choices are wider than they have ever been, and they will be wider still in 2027. Claude, GPT, Gemini, Llama, Mistral, Kimi, DeepSeek — and behind them, an alphabet of fine-tunes, distilled variants, and domain-specific checkpoints.
Do not pin the brain to a single provider. Pin the brain to a router — an AI gateway that sits between your application and the model providers, lets you swap models per request, fails over when one provider has a bad day, and gives you one bill for the lot. The router is the difference between an agent you can defend and an agent you have to babysit.
Useful ratio: the model is around 30% of the agent's quality. The tools, the memory, and the prompt scaffold are the other 70%. People who argue about which model is "smarter" by 3% on a benchmark are arguing about the wrong number.
The hands are the things the agent can actually do. Send an email. Update a calendar. Create a ticket. Push a commit. Query a database. Post a message. File an expense. Read a chart.
The single biggest mistake in agent design is giving the agent too few hands. A "customer service AI" that can read but not write is a research assistant. A customer service AI that can read, write to the PMS, send the confirmation, trigger the recall, and update the chart is a front-desk hire.
The pattern that wins in 2026 is the toolkit aggregator — one integration that gives you access to a thousand toolkits (Gmail, Slack, Linear, Notion, GitHub, HubSpot, Salesforce, every PMS on the dominant list). One OAuth handshake. One refresh-token story. One rate-limit policy. The aggregator is the leverage. Build the agent against the aggregator, not against each tool individually.
The memory is the part the agent keeps. Three kinds:
The mistake is treating memory as a feature. Memory is infrastructure. It is a system that the agent reads from and writes to on every turn, with a defined retention policy, a defined scope, and a defined access list. The agent that has memory is the agent you stop re-explaining things to.
The soul is the part that is hardest to defend in a spec but easiest to feel in a product. It is the personality file. The way the agent writes the email. The tone of voice on the phone. The brand-fit of the SMS. The escalation threshold — when does this agent hand off to a human, and how does it frame the handoff?
Version-control the soul. Put it in a soul.md file. Commit it. Review it. Roll it back when a change breaks the brand. The soul is the difference between an AI that sounds like every other AI and an AI that sounds like your company.
Because every layer is swappable. You can ship a better brain next quarter without rebuilding the body. You can give the agent new hands without retraining the brain. You can write a new soul for a new vertical without re-engineering the hands. You can store memory in a new database without changing the body's UI.
The model also tells you, on day one, what is missing. No body? It's a CLI. No brain? It's a chatbot. No hands? It's a search box. No memory? It's a stranger every call. No soul? It's a robot.
When you can name which layer is broken, you can fix it. That is the whole point of the anatomy.
An agent with all five layers is not a chatbot. It is an employee. Hire yours deliberately.
CTO