What Is an AI Agent? An 8-Agent Team in a One-Person Company

An AI agent is software that gives a language model tools and a goal, then lets the model plan its own steps, check the results and keep looping until the job is done. A chat model writes an answer. An agent opens a file, calls a system, reads what came back and decides the next step. This guide covers the definition, the working loop, and how an 8-agent team runs day to day inside a one-person company.

At the end there is a starter checklist you can copy. It begins with two roles. You grow the team later.

What is an AI agent?

An AI agent has four parts:

• Model: the language model that makes the decisions.
• Tools: real actions such as reading a file or running a command.
• Goal: a precise description of what “done” looks like.
• Loop: the repetition that goes from action to result and from result to a new decision.

A model on its own produces text. Give it tools and the text turns into action. The result of each action flows back to the model, and the model decides its next move based on it. That feedback loop is what makes an agent an agent.

Anthropic’s definition separates two structures. In a workflow, models and tools are orchestrated through code paths written in advance. In an agent, the model dynamically directs its own process and tool usage. If the path is drawn beforehand, what you hold is a workflow. If the model picks the path, it is an agent.

On the Day 60 livestream I defined an agent like this: you give a language model tools, and it “uses those tools and turns into an agent”. That is why I call the entities in AgentSpace agents. They can use tools such as team management and roles, and memory sits among those tools. In the words from the stream: “because they can use these, they can go out on real-world tasks.”

A single long prompt is not an agent. Without tools there is no action. Without a completion condition there is no loop. Two questions are enough to test any system: does it use tools, and does it read the result and change direction?

The difference between a chatbot and an AI agent

Putting three structures side by side shows who decides the steps in each one.

The boundaries overlap, and you can see all three in a single product. In ChatFlow, the WhatsApp commerce automation I run for e-commerce brands, the order pipeline runs on n8n workflows, the Shopify Admin API and a WhatsApp messaging layer. That part is a workflow. The free-form question a shopper types into WhatsApp is handled by the AI sales agent.

Workflows win on predictable work. If the same steps run in the same order every time, you do not need an agent. When the steps change with the content of the job, an agent earns its place.

How does an AI agent work?

There is no single magic component inside an agent. The same loop repeats on every task, and it reads easily as four steps.

Goal and plan

The agent starts with a goal. A good goal contains its finish condition: “this behavior works in this file and the test passes.” The model splits the goal into sub-steps. The plan does not stay fixed, because the result of the first step can change it.

Tools and checking

At each step the agent picks a tool. It reads a file or runs a command. The result comes back. The agent compares the output with the goal. If it matches, it moves on. If it does not, it tries the same step by another route.

A loop without a check step is just a sequence of commands. That is why the done-criterion has to be written down: the agent knows what to look at.

A small example: the goal is “phone validation in the form works”. The agent first reads the relevant file. Then it writes the rule and runs the test. If the test fails, it reads the output and reworks the rule. When the test passes, it reports the result.

Memory

A model carries context inside a session. When the session closes, the context is gone. Persistent memory stores decisions and preferences across sessions. In AgentSpace, agents remember decisions and preferences between sessions, and the rules that enter memory are approved by hand. The details are in giving AI agents persistent memory.

Permission boundary

Every agent that uses tools works inside an authority zone. Which folder it can write to and which commands it can run are defined up front. The narrower the authority, the more predictable the agent’s blast radius. The rules section below turns this into concrete practice, and scoping agent permissions covers it in depth.

From one agent to a team: multi-agent systems

One agent is enough for small tasks. As a task grows, a single agent’s context swells. Backend and interface work pile into the same conversation, and tests get added on top. The model has a harder time telling what to focus on. The case for splitting that load is in stop using one giant prompt.

A multi-agent system splits that load. A lead agent takes the task and breaks it into subtasks. Specialist workers own the pieces. Delegation in AgentSpace works this way: the lead splits the task, and workers run in parallel and in isolation.

There are three concrete gains:

• Clean context: each worker carries only the information for its own task.
• Parallelism: independent pieces progress at the same time.
• Isolated work: one worker’s changes do not touch another worker’s files.

The lead’s job is coordination. The lead does not write the work itself. It defines the task, hands it to the right specialist and reviews what comes back. That separation keeps the team manageable, and the same structure holds as the agent count grows.

Clear role boundaries matter as much as the number of agents. Each worker should have exactly one specialty. When the boundary blurs, two agents step into the same job.

An 8-agent team in a one-person company

I run this structure in a one-person company. One team has 8 agents: 1 lead and 7 specialists. Here is the frame.

Roles

From task file to result file

The team’s working loop rests on two files.

1. The lead writes a task file. Scope, to-do items and the done-criterion sit in separate sections.
2. The lead hands the task to the right agent.
3. When the worker finishes, it leaves a result file: what it did, and the evidence.
4. The lead reviews the result. It closes the task or writes a new one.

Because the loop is written down, information does not get lost inside a chat. Every task has a beginning and an end, and you can follow the team by reading files.

Two rules

The team runs on two rules:

• Two agents never edit the same file at the same time.
• Nobody writes “done” without evidence.

The first rule closes off collisions from the start. The second makes the lead’s review concrete. Evidence is a test output or a command result.

Three moments from the livestreams

On the Day 55 livestream I explained AgentSpace’s purpose: you set up a company for yourself, acquire AI employees, manage them under a defined system and have them build a product.

On Day 55 I also showed that the ChatFlow development work comes out of AgentSpace, and that I run it on autopilot. I had done the planning ahead of the streams. On the Day 55 stream we started the tasks on autopilot and the team worked through them in order. The planning side of autopilot is covered in autopilot sprint planning for an agent team.

On the Day 46 livestream I showed different teams working at the same time. In one team I build the AgentSpace app. In another I build ChatFlow. I go to a team and talk to its lead. Most of the time I do not even need to talk, because they run the tasks on autopilot. In the Day 46 stream I put it this way: “What we need to do is define the tasks.”

The Day 60 definition summarizes this section too: what makes an agent an agent is the tools it holds. AgentSpace gathers those tools in one virtual office. Each agent works in its own terminal and has a persistent identity.

Being small is an advantage here. One founder defines tasks. The team executes in parallel. Coordination rests on files, not meetings.

How do you build an AI agent?

To build an AI agent, you first pick a single job. No grand vision is needed. The job needs a clear start and end, and its proof must be visible. “Write the validation rule for this form and run its test” is a good first task.

Step-by-step setup

1. Choose a coding engine. AgentSpace works with 11 coding engines, and Claude Code is one of them.
2. Give the agent a role. The role should fit in one sentence.
3. Define its tools. Write which folders it can write to and which commands it can run.
4. Put the done-criterion into the task file.
5. Prepare the result file template. The agent leaves its evidence there.
6. Set the memory rules and give approval by hand.

Code and files stay on disk. You use your own API key. This setup lets you manage the agent like a local employee.

A copyable task and result template

The block below is the skeleton of a task file and a result file. Fill it in for your own work.

TASK FILE
Scope: (one sentence, which file or module)
To do: (short numbered list)
Done criterion: (which test passes, what you will see)
Files not to touch: (list)

RESULT FILE
What I did: (short summary)
Evidence: (test output or command result)
Changed files: (list)

A starter checklist with two roles

Do not jump to 8 agents on day one. Start with a lead and one specialist.

• Pick a single job and write its done-criterion
• Define a lead role (writes tasks, does not do the work)
• Define a specialist role (one area of expertise)
• Save the task file template
• Save the result file template
• Limit the folders the specialist can write to
• Give the first task and read the result file
• Do not close the task without evidence
• When the specialist works well, add a second specialist
• Write only approved decisions to memory

Once the two-role setup is stable, add a third specialist. Keep the same loop with every addition.

Ideas for a first task

The first task should be small and verifiable. A few examples:

• Update the text layout on a single page
• Add a test to a script’s output
• Scan a list of sources and produce a short summary

All three have a clear done-criterion, and the evidence shows up in the result file. Once the first task is stable, widen the scope.

Rules for running an agent safely

Because an agent uses real tools, the rules need to be written from the start. Five rules give a sufficient frame.

Permission

An agent receives only the authority its task requires. Write access is limited to one folder. The command list can change from task to task. Details are in scoping agent permissions.

Evidence

Every task closes with evidence: a test output or a command result. The evidence lives in the result file. The lead closes a task only after reading it.

File ownership

A file has a single owner at any moment. Two agents never edit the same file at once. The task file states explicitly which agent writes to which file. That keeps parallel work clean.

Human approval

Rules that enter persistent memory are approved by hand. Changes that go live pass through your approval as well. The agent proposes. The decision stays with you.

Small start

Try a new type of task first in a narrow scope. Read the result file and check the evidence. When the pattern is stable, widen the scope. This rhythm raises both trust and speed.

A task board makes tracking easier. In AgentSpace the cards work with drag and drop. A card shows a title and a description, and you assign the card to an agent.

Frequently asked questions

What is an AI agent and how do you build one?

An AI agent is a language model that uses tools and works in a loop toward a goal. Building one takes a role, a tool set and a written done-criterion. You start with one job and two roles, and you put evidence requirements and permission limits in place from the beginning.

What are AI agents used for?

They take on work whose steps are not known in advance. Software development and research are typical examples. Because an agent can use tools, it can go out on real-world tasks. For fixed, repeating work, workflow automation is the better fit.

How do you use an AI agent?

You give the agent a task file. Scope, to-do items and the done-criterion sit in separate sections. The agent does the work and leaves its evidence in a result file. You review the result, then close the task or write a new one.

What does “agentic AI” mean?

Agentic AI is an approach that does not finish a job with a single answer and instead moves step by step. The model takes a goal and acts through tools. It reads the result of every step and updates its plan.

What are examples of AI agents?

A development agent that writes code and runs its tests is one example. A research agent that scans a source and produces a summary is another. A sales assistant that answers shopper questions on WhatsApp belongs to the same group. What they share is tool use and a goal-driven loop.

What is the difference between an AI agent and an automation workflow?

In a workflow, you draw the path in advance and the steps run in a fixed order. In an agent, the model chooses the path and the tool order while it runs. Predictable, repeating jobs suit workflows. Jobs where the steps depend on the content suit agents.

What is a multi-agent system?

A multi-agent system is a setup in which several agents share roles. Usually a lead splits the task and specialist workers carry out the pieces in parallel. Clean context and isolated work are the main gains of this structure.

Related posts

• My AI Agent Team Runs Without Me: 3 Lessons From Autopilot
• Stop Using One Giant Prompt: How I Run AI as an 8-Agent Team