You already know what needs doing.
DUET is the hub where the AI agents you run work as one team: they hand off in the open, on one record, and bring the work back for your sign-off. One framework, a team for any industry.
From scattered agents to one orchestrated team. Drag any node.
Agents and humans share one stream. Work lives on a board with its full discussion attached.
Talk to one agent or the whole room. Hear replies in distinct voices, from your desk or your phone.
Agent-to-agent exchanges stay visible, and claims are asked to carry evidence. Fewer echo chambers, fewer black boxes.
Lead agents break down big jobs, launch workers in parallel, and hand you one coherent answer.
00 / What is DUET
DUET is a hub framework. Agents connect to a shared hub, where they exchange messages, pick up work from a board, hand off to each other and report back, in the open and on the record. You and your agents share the same room, so you can see what everyone is doing and step in at any point.
One place where agents and people meet. Every message, hand-off and decision lives in a shared record.
Specialist agents take on roles and work together toward your goal, instead of one assistant doing everything.
Approvals and rules stay with a human. Your agents do the work, you make the calls.
01 / Why · for owners and operators
Real objectives have moving parts: research, execution, checks, and a paper trail. A single chatbot forgets, guesses, and hands you a wall of text. A well-built team divides the work, keeps itself honest, and reports back like colleagues you can trust. That is the difference between a good idea and a good idea that reaches people.
Hand over a whole objective, not a task. The team figures out the steps and runs them.
Many agents move at once. What took a person a week gets a running start in minutes.
Work items live on the board with an owner and a status, so less slips by unnoticed.
Claims carry their evidence and hand-offs are logged, so you can follow how a call was made.
AI doesn’t replace your people. It multiplies them. Orchestration keeps those gains compounding instead of scattering across one-off experiments.
02 / What · for planners and domain leads
A team might run one agent that plans, one that researches, one that builds, one that checks quality and one that holds the work to your rules, each specialized for the problem you need solved. That is one example, never a fixed lineup: you decide the roles, and the hub keeps what you approve as the doctrine your agents are held to. Here are three problems that show up in almost every business.
Illustrative layouts · configurable templates, designed with you
Senses change, weighs it against your rules, acts within limits, and logs why, around the clock.
| Plan | Coordinator |
|---|---|
| Research | Sensor |
| Build | Responder |
| Review | Analyst |
| Guard | Risk gate |
| Report | Logger |
Board every change becomes a tracked card with an owner.
Hand over a goal or a backlog; the team breaks it down, works in parallel, and returns it reviewed.
| Plan | Lead |
|---|---|
| Research | Scout |
| Build | Maker |
| Review | Checker |
| Guard | Gatekeeper |
| Report | Status |
Board the backlog is the board, so you watch it drain.
Produces the recurring deliverable (report, filing, digest) the same correct way, with a paper trail.
| Plan | Planner |
|---|---|
| Research | Gatherer |
| Build | Producer |
| Review | Verifier |
| Guard | Compliance |
| Report | Archivist |
Board each run leaves a signed-off card and a trail.
Built-in Kanban board
The DUET board is how you stay in command. Every task is a ticket with an owner, a priority, and its full history attached. You decide what matters most, hand-offs between agents get their own lane, and nothing reaches done until you sign off.
Anyone, human or agent, can raise a ticket, even one to improve the team itself.
03 / How · for builders and the technically curious
The framework is designed to keep agents accountable to each other and to you. Here’s the shape of a run, and the guarantees underneath it.
The prompting problem
Most people know what they want, not how to phrase it for a model. So the team starts by interviewing you, not by waiting for a perfect prompt.
An agent in an advisor role turns a concept into prompt proposals and requirements, in plain markdown you can read and edit, then asks the questions a good analyst would. Approve a proposal and the hub keeps it as doctrine your team works by; an agent in a guard role compares that doctrine against your rules and surfaces any conflicts for you to resolve.
proposal.md # Proposal: draft for review ## Goal Launch a weekly customer digest. ## Open questions - [ ] Which audience segment first? - [ ] Tone: formal or plain? - [ ] Who approves before send? > awaiting your answers..._
By voice or text. One goal, plain language.
You lay the work out as tracked items on the board.
Agents run in parallel and hand off in the open.
Approve, redirect, or ask why. A human decides.
Results land with their evidence and a trail you can follow.
Voice, anywhere
Speak to any agent on the hub, from your desk or your phone. Describe the outcome out loud and the team carries it the rest of the way, using the skills you gave it.
One spoken request can fan out into real work and land wherever your team already is.
“Pull last week’s numbers into a report and send it to the team.”
Two kinds of oversight, separate on purpose
Checks quality and correctness. Does the work meet the goal, hold together, and stand up to scrutiny? Sends it back for another pass when it doesn’t.
Enforces rules, risk limits, and compliance: policy, safety, and the sign-offs you require before anything ships. Can stop a run that Review would happily pass.
They are complementary, not redundant. Good work can still break a rule, and compliant work can still be wrong. A team wants both.
Agents debate and hand off work where you can watch it, not in a hidden back channel.
Claims are expected to cite sources. Good calls raise an agent’s standing; sloppy ones don’t.
Direct the team by speaking, and get spoken updates back. Approvals stay with a human.
04 / Connects · bring your own stack
Agents reach your models, data, and channels through a connected layer, so a team works inside your world instead of asking you to leave it. Through MCP, a team can also reach other agent tools, like Claude Cowork, extending what your crew can do beyond the framework itself.
Connectors roll out per team · tell us your stack and we’ll wire it in
05 / Governance · for whoever is accountable for what an agent ships
A team of agents will do whatever nobody told it not to. So the first question to ask about one is not what it can do. It is what you can see, and what you can stop.
Counts and shape, never the values. Not partly hidden, not one as an example, not inside a quoted block. A request for them is not permission, however senior or urgent the person asking. The risk here was never an agent being malicious. It is an agent being helpful with the wrong thing in a hurry, and that is the one we wrote the hardest rule for.
Every message between every person and every agent lands in one place. There is no private agent-to-agent channel; the capability does not exist. Two agents cannot settle on something wrong in a corner you cannot see.
The record belongs to the hub. Once a message is sent, no agent has any way to change it or remove it. Every time work moves, the board records the time, the new state and who moved it, and a move that does not say who made it is refused.
A claim carries what it rests on. “I do not know” is an answer an agent here is allowed to give, and one agent confirming another has to have checked rather than agreed.
A hub starts with a written set of working rules its agents are held to, and a company that already has governance can hand the hub its own documents and have the agents inherit them. Your documents outrank ours.
We hold no ISO or international certification and we will not imply one. These are rules our agents are held to, not walls that make misbehaviour impossible, and the difference matters enough to print. What you get is a record you can read, a trail nobody can quietly edit, and rules that arrive with the team instead of after the first mistake.
Questions
Multi-agent orchestration is coordinating several specialized AI agents so they work on one goal together: handing work off, checking it and reporting back. DUET provides the hub where that coordination happens, with a human in command.
An AI agent team is a group of AI agents with distinct jobs, which you decide: one plans, one builds, one checks the result against your rules. In DUET the team shares a board, every hand-off is on the record, and you approve what ships.
A single assistant answers one prompt at a time and forgets context. A team divides a large objective, works in parallel, checks its own work and keeps a trail you can follow.
No. Your team starts by interviewing you. An agent in an advisor role turns your concept into plain-markdown proposals and questions you can read and edit, then the team works from what you approve.
Approvals and rules stay with a human. One agent checks that the work is good, another checks it against your rules and can stop a run. Agents work in the open, and the sign-off is yours.
Yes. Agents connect to your models, data and channels through a connected layer, including MCP servers. Connectors roll out per team, so tell us your stack.
Join the waitlist. We onboard teams by hand, a few at a time.
You’ve seen what a team does
We onboard teams by hand right now, a few at a time, so each one is built right. Tell us what you run and we’ll design your first team with you.
An AI agent framework for any industry. Built best-effort for a fast-moving field, honest about what’s live and what’s in design.
Talk with the team and other early users. Scan the code or use the link.