How multiple AI teammates work together
A guide to multi-agent work on Nagent today: people and AI teammates in one shared context, an AI chief of staff who routes the work, handoffs inside a team, and approvals that decide what runs without a person.
Browse documentation
Who this is for
You run growth work, in marketing, sales or content, and you want more than one AI agent on it without ending up with a pile of tools that do not talk to each other. Or you are the person in your company who sets Nagent up and wants to know how the agents share work before you trust them with any.
This page explains the shape of multi-agent work on Nagent. For the screens themselves, start with the team room and anatomy of a team.
The short version
- Work happens in a team room, where your people and your AI teammates read the same brief, the same record and the same board.
- Each department has an AI chief of staff: MIRA for Marketing, SERA for Sales, CREA for Content. The AI chief of staff speaks for its team and routes work to the team lead who owns it.
- A lead can hand off to a teammate on its own roster. One message pulls in at most three more teammates, and the teammate's own level decides what it may do.
- Every action an agent can take is checked against its level before it runs. Anything it is not yet trusted with waits for a person.
- Every agent has a budget, and a paid call is checked against it before it is made.
Departments, leads and teammates
A department owns one outcome. Sales owns qualified meetings, Marketing owns reach and pipeline cost, Content owns output and the brand line.
Each department installs a team into your workspace: a lead and the teammates who work under it, seated in a team room. Today that means:
- Marketing, led by MIRA, with CREA and RUPA for content, DRIS for AI search visibility and NIA for Google Ads. HOOK (ad creative, Beta) and Echo (reputation, Beta) report to MIRA on the org chart.
- Sales, led by SERA as AI chief of staff, with NORA, DEXA and RIVA under her and LIPI and VANI writing e-mail. DEXA and RIVA join when your workspace has them.
- Content, led by CREA, with RUPA for images and the FOLIO teammates for long-form publications.
The org chart on the Agent Workbench shows who reports to whom. A reporting line is a governance fact: who speaks for a team is read from one place, and a team with no lead refuses to act rather than electing one.
Briefing a team
You brief the lead with the outcome, not a task list. In the room:
- Type @ to address someone. Addressing an agent asks it to answer; addressing a person only records the message.
- The room reads what kind of message it is: work an agent can do now, an instruction nothing can run yet, or a question.
- Work an agent can do now starts at once, after the agent states its assumption.
- An instruction becomes a card, "Instruction read, waiting on you", with Confirm, Edit and Not a task. Only a person can confirm it.
- A confirmed task lands on the board, attributed to the person, and the agent named on it picks it up and says so in the thread.
- A question gets an answer in a thread, grounded on the team's record and Knowledge Hub, with sources you can open.
- The result comes back as a file. Accept puts it on the record; Send back returns it with a reason.
Name the agent when you brief. A task with nobody named is not picked up.
How work moves between teammates
A lead routes; the teammate who owns a piece of work does it, and keeps its approvals, its budget and its record.
Handoffs are bounded on purpose:
- A lead may hand work only to a teammate on its own roster, or to an agent that reports to it.
- A teammate the lead's plan sends work to does that work; it does not plan for the team again.
- One message can pull in at most three more teammates, and never the same one twice.
- The receiving agent's own level and budget decide what it may do. A handoff never lends the lead's trust to someone else.
Delegation spreads effort. It does not launder autonomy.
An example: outbound to a new segment
Say you want first-touch e-mail to a segment you have not sold to before.
- In the Sales room you brief the lead: "Find ten accounts that fit our D2C profile, research each and draft a first e-mail."
- NORA finds accounts that fit your customer profile and researches each into a sourced dossier.
- The draft e-mail for each account is written from that dossier and your brand voice.
- Nothing is sent yet. On a new workspace every outbound send waits for a person, whatever level the seat shows. The drafts appear under "Waiting on you" in the room and on the Actions page.
- You approve, edit or send back each one. Every approval is recorded, and that record is what a later promotion is judged on.
The same shape holds for a content brief to CREA, a campaign brief to MIRA or a visibility audit with DRIS: the lead plans, the owner does the work, and a person approves anything that goes live.
Levels, approvals and earned autonomy
Agents sit on five levels, from L1 Observe to L5 Lead. Each action an agent can take declares the level it needs to run without a person; pure reads always run.
- When the agent's level meets the action's bar, the action runs.
- When it does not, an approval request is written and nothing runs.
- A paused or switched-off agent takes no action with a side effect.
- Destructive changes, and changes to how work is prioritised, ask a person every time, at every level.
- Publishing needs a person below L5. At L5 an agent, such as a Chief of staff, publishes alone.
Autonomy is earned, not deployed. Each night Nagent nominates an agent for the next level once it has spent 30 days at its level with at least 25 decided approvals and fewer than 5 in every 100 dismissed. A workspace admin approves or declines, and a record that slips drops the agent a level. See autonomy and governance.
Budgets
Every agent has a daily cap and a monthly cap. Before a paid call, Nagent checks what the agent has left under both, counting calls already running. When a cap is reached, the agent's setting decides whether further paid work queues, the agent pauses, or the breach is recorded and work goes on. Above the agents, a workspace daily limit refuses new paid work while reading, reviewing and approving still work. See credits and budgets.
Grounding
Agents answer from your workspace's own record and Knowledge Hub, with sources, and say so when the record does not hold the answer. A confident wrong answer costs a team more than an honest miss. Add what your teammates should know in the Knowledge Hub.
Workflows
A workflow is a job the team repeats, written down as steps that agents do: a weekly report, or a reply to every new lead. A Team lead or Chief of staff builds and changes the team's workflows, or an Organisation admin. Members run them from the room, or by asking N.
Each step keeps its agent's own level, budget and approvals. A step that needs a person waits, and the room's card says Waiting for approval. After a day the room gets one reminder. A request nobody answers expires, after 7 days for most, and the run ends unless the workflow has a path for that. See building a workflow.
Building your own team
The prebuilt teams cover Marketing, Sales and Content. When you need a team shaped to your own work, you can build one in the admin: agents, a lead they report to, the tools and skills they need, a room and a recurring workflow. See building a team of your own.
Checklist before you rely on a team
- The team has a lead, and every brief names the agent who owns the work.
- The Knowledge Hub holds what the team needs to answer from.
- Levels are set so anything irreversible waits for a person.
- Budgets are set per agent, not left on the default.
- Someone on your team owns the "Waiting on you" queue every day.