AI agent architecture explained, part by part
AI agent architecture is the set of parts an AI agent is made of and the way they work together. An AI agent is an AI that does not only answer questions but also carries out tasks, such as updating a record or drafting an email. It has six main parts: instructions, a model, tools, context, limits and a log.
Why a business owner should know the parts
You do not need to build an agent to benefit from this. Knowing the parts helps you ask good questions when you compare products. It also helps you see where to look when an agent gives a poor result. The cause is rarely "the AI". It is usually one specific part.
The six parts, compared to a new employee
| Part | What it is | For an employee, this is |
|---|---|---|
| Instructions | A standing text that sets the job, the rules and the tone | The job description |
| Model | The AI that reads and writes text | The person's ability to think |
| Tools | Actions it can take, such as "search records" | The phone and the programs on the desk |
| Context | The request and the information it is given right now | The papers on the desk today |
| Limits | What it may see, what it may change and what needs a yes | Access rights and sign-off rules |
| Log | A record of what it did | The file you can check later |
The model is only one of the six. A clever employee without a job description, without access to the files and without rules would not get far either.
One request, followed through the parts
A project lead writes: "The timber delivery is a week late. Move every task that depends on it."
- The agent gets its context. It receives its instructions, the request and the open project.
- The model picks a tool. It does not answer yet. First it asks for a search: which tasks depend on the timber delivery?
- The limits are checked. The search runs with the project lead's access rights. It finds six tasks.
- The model proposes a change. It suggests new dates for the six tasks.
- A person confirms. Changing several records needs a yes. The project lead sees the old and new dates and agrees.
- The result is logged. The dates are changed, the agent writes a short summary, and the whole run is saved.
One point is worth remembering. The model never touches your data directly. It only asks, and ordinary software decides whether the request is allowed. That is why limits can be relied on. They do not depend on the AI behaving well.
Which part to look at when something goes wrong
- The answer is wrong or made up. Look at the context. The right information was probably missing.
- It did the wrong thing. Look at the instructions and the tools. Unclear wording confuses an AI just as it confuses a person.
- It did something it should not have. That is a gap in the limits. Better wording in the instructions will not fix it.
- Nobody can say what happened. The log is missing or too thin.
Small agents and large ones share this plan
An agent that sorts incoming emails and an agent that plans a product launch are built from the same six parts. The larger one simply has more tools and more information to work with.
How Worknano handles this
In Worknano these parts are fields you can open and edit. An agent has a name and description, instructions, tools, a model and a list of other agents it may hand work to. The limits come from the workspace: an agent can only read what the signed-in user may read, and it makes changes through a preview that a person confirms. A log keeps every agent run, with what went in, what came out and which model was used.
- AI agents in Worknano: what an agent is and where it runs
- How to edit an agent in Config
- Build your own agent
Create a free workspace and open one of the ready-made agents to see its parts.