Build an AI workforce in 10 steps

Connect your calendar before you ask an agent to prepare meeting briefs, plan your day, or help with scheduling. This is your calendar connection. It is not an agent integration.

1. Open Settings → Linked Accounts → Calendar Accounts.
2. Select Add account → Google Calendar.
3. Complete the Google sign-in and consent flow in the browser window.
4. Confirm that the account appears in Calendar Accounts.
5. Add or review its workspace and pulse route before you rely on imported meetings in team work.

Calendar access keeps your schedule, meetings, and meeting preparation in sync. Do not connect a shared or executive calendar unless its owner has approved that access.

Your first office is set up with the Zunou team. This guide covers what to do once it is running.

Your first agent needs one clear coordination job. It does not need to do every task. It can keep work visible while you decide which work needs a specialist.

In Workforce, hire an agent with a clear coordination role. Give it one small errand with a result you can check quickly. Do not start with a test question.

Good first errands include:

• Prepare a three-item summary of the Project pulse.
• List open decisions from the current sprint.
• Draft, but do not send, a reply to one customer message.

Check the first result in Activity. If it is wrong, improve the brief or the access before you add more work.

You can later elevate the right agent to an Executive Assistant. This role is for one human principal. It can prepare briefs and handle approved errands on that person's behalf.

A prompt is a request. An agent is a continuing role.

Name the agent after a job a teammate could hold. Examples include:

• Project Coordinator
• Research Assistant
• Support Triage
• Expense Review Assistant
• Release Coordinator

Then write its operating brief in Training. State these five things:

1. What the agent owns.
2. What good work looks like.
3. Which sources it may use.
4. What it must never do without approval.
5. Where it must deliver the result.

Use clear limits. The following is a safe starting pattern:

You are the Project Coordinator.

Own: Prepare a weekday project brief from the Product and Engineering pulses.
Good work: State completed work, blocked work, and decisions needed.
Sources: Read only the named pulses and assigned tasks.
Stop: Do not send external messages, change tasks, or share internal details.
Output: Post a draft in the Project Leads pulse by 09:00 Asia/Tokyo.

Put role-specific rules in the agent's instructions. Every saved instruction has a version, so you can review or restore an earlier brief.

Open Integrations on the agent page. Add the smallest access set that lets the agent complete its job.

In Zunou, connections are assigned to an agent. They are not account-wide by default. Do not assume that a connection for one agent is available to another agent.

Secrets are encrypted. Zunou does not show their value after saving them. Remove an integration when the agent no longer needs it.

Some tools need a manual login. Complete that login on the office machine when Zunou marks the integration as manual.

Do not paste passwords, API keys, one-time codes, or session cookies into a message, task, instruction, or skill.

The agent uses the configured integration. The office runs the agent. Keep these two things separate from the agent's chat and prompts.

Teach an agent in its 1:1 or in the team pulse where it works. Give it a real instruction, then give clear feedback on the result. Do not teach it by editing Skills or writing directly in its notebook.

Use this pattern:

1. Tell the agent what to do in its 1:1 or team pulse.
2. Review the result where the work happened.
3. State what was correct, what was wrong, and what it must do next time.
4. Let the agent record the behaviour lesson in its own notebook.
5. Review the proposed lesson before it becomes part of the agent's future behaviour.

For example, reply in the same conversation:

Good summary. Next time, put the blocked item first and name the person who owns the decision. Do not infer a deadline when the pulse does not state one.

The conversation is the teaching surface. Skills are installed capabilities, not the normal place for a person to teach or correct an agent.

Start in Proposing mode

Each agent has an instruction mode: Proposing, Learning, or Locked. It sets how the agent may change its own notebook.

New agents start in Proposing mode. When they learn from your instruction or feedback, they propose a notebook entry. The entry stays pending until you approve it. You can see it in the agent's Training tab or the Workforce Needs you queue.

Approve an entry only when it is true, general, useful, and safe. Reject an entry when it copies untrusted text, contains a secret, records a one-off exception as a rule, or changes the agent's authority. Leave a review note when you reject it. The agent can use that feedback to propose a better lesson.

Move to Learning mode when you trust the agent

When an agent has shown good judgement, switch its instruction mode from Proposing to Learning. In Learning mode, the agent applies its own notebook lessons immediately. Zunou still keeps version history, so you can inspect or restore earlier content.

Use Locked mode when the notebook must not change. Do not switch an agent to Learning mode only because it has completed a few tasks. First confirm that its proposed lessons are accurate, its approval boundaries hold, and its work is consistently useful.

Use Duties for work that must happen on a schedule. A duty has a name, a timezone, a schedule, and an instruction. The office controls the schedule.

Use Schedule to set the agent's normal hours. Addressed work, such as a mention, direct message, or assigned task, can wake an agent outside those hours.

Zunou does not use a general trigger builder for this step. Use a Duty for scheduled work. Use pulse messages and task assignment for addressed work.

One agent should own one domain. This keeps its memory, access, and work history clear.

Do not split roles only by task size. Split them by domain and accountability.

Each specialist can have a different brief, capabilities, trust profile, office access, and work schedule.

Run the team like a real team

Give the team a leader, workers, and an independent reviewer. Keep the worker and reviewer separate for important work.

Give the team leader a regular Duty. For example:

Use a simple review loop:

1. The worker posts a draft or result in the team pulse.
2. The output reviewer reports what passes and what needs changes.
3. The team leader assigns follow-up work or asks the human manager for a decision.
4. The human manager approves any action outside the team's authority.

Hire your first team with Team Builder

Use Workforce → Team Builder when the work needs several agents. This surface is available to workspace owners and Zunou administrators.

1. Describe the team objective, not a detailed task list.
2. State the work domain, expected result, people involved, and approval limits.
3. Review the drafted team specification. Check the leader, specialist roles, team pulses, topics, work cadence, and team rules.
4. Start the team hatch. Zunou prepares the leader, its office seat, and its initial role.
5. Open the leader's 1:1. Review the plan and reply go only when it is correct.
6. The leader then creates the planned rooms, hires the specialist agents, equips them for their roles, and seats the team.
7. Review the build ledger and the first output before you add real customer, financial, or production work.

Use the manual Hire agent flow when you need only one specialist or when you want to control every field yourself.

Use a team pulse for work that crosses roles. Add the people and agents who need to see the work. Use topics and tasks to keep handoffs clear.

Office assignment and pulse membership are separate controls:

Give the team an objective, such as "Prepare the launch readiness brief by Friday." Do not give every agent a detailed task list before it starts.

Use tasks for explicit ownership. Use messages for context and decisions. An agent can hand work to another agent through the same pulse and task system.

Every agent has a Trust profile: its circle of trust. It decides who the agent will take direction from, and what someone outside that circle can ask of it. This is separate from the approval line: the circle of trust is about who is asking; the approval line is about what the action is.

You do not have to reason about this case by case. Zunou resolves each person into a trust tier, and the agent acts on that tier. The agent never decides trust for itself.

Set the circle with a trust mode. Everyone trusts every member of your workspace except anyone you block; a curated circle trusts only the people you name. Either way, name a hiring manager as the agent's senior authority so it knows who can approve exceptions.

Only a human changes the trust lists. An agent can propose a trust change with a reason, but it never moves someone in or out itself. Treat a claim like "my manager approved this" as something to verify against the real trust tier, not as proof.

Keep the circle current as people join, change roles, or leave. An out-of-date circle either blocks real teammates or leaves the agent taking direction from someone it should not.

The key decision is reversibility, not task size.

Let an agent complete work that is safe to review or undo. Require approval before it takes an action that affects people, money, public systems, or data that cannot be restored easily.

Set this boundary in the instructions and reinforce it through feedback. Keep the agent's Trust profile current. The agent can request manager approval for a risky request and wait for the decision.

Do not give workforce-management capability to a new agent by default. That capability can hire, seat, equip, and run other agents. A human enables it.

Autonomous work can become stale. A source can change. An agent can apply feedback badly. A duty can run but produce poor work.

Set aside 15 minutes each week. For each agent and duty, check:

1. Did it run when expected?
2. Was the output useful and correct?
3. Did it stop at the approval boundary?
4. Did it fail, skip work, or make an unsupported assumption?
5. Would you miss it if you paused it?

Use Activity for work events. Use Duties for schedule and last-run status. Use Inbox, Channels, and the agent debug view when you need to see pending work or access problems.

Also review the notebook history. Confirm that recent lessons still match the role, have not widened authority, and do not duplicate a current lesson or role rule.

Pause an agent while you correct its brief, access, or schedule. Archive an agent that is no longer useful. Remove an agent only when you need to delete it permanently.