Part 1 · Lesson 4 of 9 · 3:22 video

Three types of AI agents, one brain

Choose what starts a task: a person, a trigger or an assistant.

Course content Lesson 4 of 9

Lesson notes

The ideas from the video, with examples.

What starts the work

	Chat	Workflow	Your own assistant
Started by	a person asking	a schedule, a webhook, a manual run or a request in chat	the assistant a person already uses
At Wren & Page	Sam asks how each branch did	Nora's report at 8:00	Maya asks from her assistant
Who runs the model	the hub	the hub, only in the steps that need it	the assistant, with its own model
In Hubzoid	web chat, Slack, WhatsApp, Telegram	`workflows/` in Python, `schedule/` in plain English	the hub's MCP server
Built in	lesson 6	lesson 7	lesson 8

One brain, three ways in
All three read the same files. Change the returns policy once, and every way in uses the new version.
- wren-page/one hub, one brain
  AGENTS.mdinstructions and starter questions for chat
  knowledge/shared by all three
  skills/shared by all three
  tools_local/shared by all three
  workflows/code that starts on a schedule, a webhook or a request
  schedule/plain-English tasks, on a schedule or an event
  .envsettings, including the MCP server for the third way in

What is shared, and what is not

Shared across all three	Kept separate
Knowledge, skills and tools	Each person's conversation history
The agent's instructions in `AGENTS.md`	Each workflow run and its saved steps
The rules for restricted tools	Each person's permissions

So "one brain" means one source of truth, not one shared memory of every conversation.

Exercise

Try it: sort your own tasks

List five tasks your team repeats.
For each, ask what starts it: a person, a clock or event, or their own assistant.
Mark the one to build first. The most frequent usually wins.

task-sorter.md

| Task                          | What starts it        | Way in    |
| ----------------------------- | --------------------- | --------- |
| How did a branch do yesterday | a person asks         | chat      |
| Morning sales report          | 8:00 every day        | workflow  |
| Check a publisher invoice     | Maya, in her own tool | assistant |

If you're stuck

My task fits two ways

That's normal. Start with the way it happens most often, and add the other later on the same brain.

Can a chat message start a workflow?

Yes, when the workflow offers it. A code workflow declared with as_tool=True becomes a chat tool that runs as the person asking, and schedulable=True lets people save their own schedule for it. Both need a grant from an administrator. A workflow also starts on its schedule, from a webhook, or by hand with hubzoid schedule run.

Still stuck? Run hubzoid doctor on your hub, then search GitHub issues (opens in a new tab) with your version and the full error. Setting it up for a team? Request a demo.

Go deeper

The HubWhat a hub holds, and how the three ways reach it.
Workflows and schedulesCode workflows, plain-English tasks and triggers.
Web chatThe chat your team signs in to.
Connect an assistantReach the hub from Claude Code, Codex and other MCP clients.

Transcript and what the video shows

Intro 0:00

Welcome back to Inside AI Agents. So far, we've built the brain. Knowledge, skills and tools. This time, the three ways a team puts it to work.

Three people, three habits 0:12

Sam asks questions all day, in chat. Nora wants the morning numbers waiting at eight, without asking anyone. And Maya works in her own AI assistant, like Claude Code or Codex, and doesn't want another window.

The quick fix is three separate agents. Each one gets its own copy of the policies, the steps and the tools.

Remember when the returns policy changed to forty-five days? Now someone has to change it three times. Miss one, and Sam's chat says forty-five, while Nora's report still says sixty. The three agents slowly drift apart.

What actually differs 0:50

Look closely, and the three agents are almost the same. Same knowledge. Same skills. Same tools. What really differs is how the work starts.

A person asks 1:01

One. A person asks, in the chat they already use. The web, Slack, WhatsApp or Telegram. Sam types, how did each branch do yesterday? The harness runs its loop, and answers right there in the chat. That's a chat agent.

Something else starts it 1:15

Two. Nobody asks. Something else starts it. A schedule, like eight every morning. Or an event, like a new order arriving from another system.

The steps run in order. Most are plain code, like pulling yesterday's sales. Where a step needs judgment, like spotting anything unusual, it runs the same harness, with the same brain, just for that step. Then it saves the report, and emails Nora a link. It's in her inbox before the shop opens. That's a workflow. The same steps every time, and thinking only where it's needed.

Your own assistant 1:50

Three. Your own assistant connects. Remember MCP from last time? It brought tools into an agent. Here, it runs the other way. The brain offers its knowledge, skills and tools, through MCP. Maya's Claude Code connects, and asks about yesterday's sales, without leaving her window. Her assistant brings its own model. The brain brings what the shop knows.

One brain 2:14

So don't build three agents. Build one brain, with three ways in. Change the returns policy once, and the chat, the morning report and Maya's assistant all use it.

And the three ways overlap. Maya can ask in chat in the morning, and from her own assistant in the afternoon. Same brain, both times.

One brain, three ways in 2:35

So those are the three types of agents. Really, they're three ways into one brain. A person asks in chat. A schedule or an event starts a workflow. Or your own assistant connects through MCP.

One brain. Three ways in.

Try it, part two and next 2:52

Try this. Pick one task from your week. Ask what starts it. A person asking, a clock or an event, or your own assistant. Keep that task in mind.

That's the end of part one. In part two, we get hands on. We'll set all this up with Hubzoid, an open-source platform for your team's internal agents. Your skill and your tool go in first. Next time, we explore the brain, file by file.

What the video shows

Chat starts with a person, workflows start on demand or from a trigger, and an assistant can call the shared Hub over MCP.
They can share knowledge and tools while keeping histories, workflow state and permissions separate.

Up next: lesson 5, 3:22The brain in detail: build your first hub with HubzoidStart lesson 5