You added agents faster than you added a protocol.

Capability arrives in an afternoon. Coordination almost never does.

A research agent drafts. A copy agent writes. A support drafter answers. A second human, not seeing the first run, kicks the same worker again. Three versions circulate. Nobody can say which one shipped.

That is not a model failure first.

That is a coordination gap.

The Multi-Agent Organization names the gap and installs the protocol that closes it.

The org chart is the wrong map

An org chart shows reporting lines. Who answers to whom. Useful for accountability. Useless for the path a deliverable actually takes.

Work lines are different. The research agent feeds the copy agent, which feeds a human reviewer, which feeds the docs agent, which feeds the support drafter. The chart does not show which brief is current, where approval happens, or where a source went stale.

Humans sometimes backfill that mess. They overhear. They ask. They notice two people doing the same task.

Agent workers do none of that. They execute a slice and stop. Every coordinating instinct a human supplies for free becomes a gap you have to fill on purpose.

Reporting lines tell you who is accountable. Work lines tell you how the work moves. Only one of those maps leads you to the deliverable.

What a multi-agent organization actually is

A multi-agent organization is a team of human operators and specialized agent workers pursuing shared goals through coordinated work.

That is the whole definition. Not a lab. Not a science-fiction premise.

If a few humans and a handful of agent roles already draft, research, reply, and pull data, you are already running one. The only open question is whether you run it on purpose.

Two terms keep the spine straight:

A human operator defines the work, reviews the output, and owns the outcome. Accountability does not transfer to a tool.
An agent worker is a specialized process with bounded authority. It holds a role, not a position. A task, not ownership.

“The agent answered it” describes how accountability went missing, not what actually happened.

Once you accept the roster you already have, you stop optimizing only the humans (hire, train, manage) and start optimizing the seams: who is chartered, how work hands off, where it is reviewed, and how the team stays on one picture of reality.

This is not a swarm demo

A CrewAI crew, a LangGraph graph, or a swarm notebook is a runtime. Models call tools. Messages pass. Loops run. That can be useful.

It is not an organization.

An organization needs answers the runtime will not invent for you:

who may spend money or talk to a customer
which brief is current
what “approved” means
who owns a wrong reply
how last week’s defect changes next week’s charter

You can implement CHORUS with those runtimes. You can also implement it with a shared doc and a task board. The protocol does not care.

Buying another orchestration product and hoping the process arrives with it is the expensive version of the same mistake. Tools store state. They will happily hold three conflicting versions of your landing-page copy in three tidy cards. Approval is a decision your team has to write down.

Protocol first. Tools second. Every time.

If your coordination cannot survive a tool migration, you do not have a protocol. You have a hostage situation.

CHORUS: six skills, one loop

CHORUS is the coordination protocol. Six skills. Each one closes a specific seam.

The name is a method, not a product. If you have seen software sold under the same word, this is not that.

What Review finds, Update writes back into the charters, and Sync makes visible. The protocol improves as it runs.

Charter

Define the role before you run it. Same seven fields for a human or an agent:

Purpose. Why the role exists, as an outcome, not an activity.
Inputs. What must arrive, in what form, before work is legitimate.
Outputs. Format and shape someone can check.
Authority ceiling. What the role may do without asking, and what it may never do.
Success criteria. Checks, not feelings.
Review requirement. Who reviews, against which criteria, before downstream use.
Escalation. Where the role stops and raises a hand.

A worker without a charter improvises. Improvisation is how pricing language gets invented.

The template does not change between a human operator and an agent worker. The completeness does. A human will sometimes notice an unstated assumption and ask. An agent will not. So the charter for an agent must be more complete, not less. It is the only judgment the agent has.

Handoff

A charter is the standing contract. A handoff is the work order.

Context dies at boundaries. The receiver reconstructs what the sender knew, or, more often, substitutes something plausible. Humans ask. Agents substitute. The output arrives looking complete against a context you never provided.

Carry a context packet with six fields:

Goal. What done looks like.
Constraints. Length, tone, scope, what to avoid.
Inputs. Exact artifacts, with the current one marked current.
Provenance. Who or what produced those inputs, from what.
Acceptance criteria. The checklist the next reviewer can verify without asking you.
Deadline. When, and what depends on it.

“You know the one” is not a handoff. “It is all in the thread” is archaeology, not a work order.

Orchestrate

Parallel work needs routing, not hope.

Draw the flow:

Serial where step B needs step A’s output
Parallel where it does not
Fan-out to split independent branches
Fan-in to merge them into one coherent artifact
Escalation to a human when confidence is low, the action is irreversible, or the policy has not been written

Fan-in is synthesis, not a folder. If two branches disagree and you concatenate them, you ship the contradiction.

Five symbols are enough for the graph: box, arrow, split, merge, gate. Mark a gate E to stop and decide, R to stop and check. If you cannot draw the flow on one page, you are still holding it in one head. That does not scale past a quiet week.

Review

Unreviewed agent output is technical debt. It looks done. It bills later, usually in front of a customer.

A review gate checks written acceptance criteria. Vibe-checking is not a gate.

Match the gate to blast radius:

Automated check for mechanical properties (links, banned terms, schema)
Spot-check for medium-consequence volume
Full read for irreversible or customer-facing work, by someone who did not produce the draft

A gate that exists on the board and never actually reads is review theatre. That is worse than no gate, because the team thinks it is safe.

Update

A review that changes nothing is a review you ran for the file.

Feed findings back up a ladder:

the prompt
the examples
the tool config
the SOP or playbook
the charter

Fix it at the prompt the first time. The third time the same class of defect appears, stop patching. The prompt is where the problem shows, not where it lives.

Sync

Shared state beats status meetings.

One surface. Four status words: queued, running, blocked, in review. One short ritual that asks what shipped, what the gate caught, and what you are changing because of it.

If the surface is not the place decisions land, people stop reading it. Then it rots. That is sync debt.

Install the triangle, not the binder

Do not roll out six skills in a week. That produces a document nobody follows.

The ninety-day triangle from the book:

Days 1 to 30. Charter every role, starting with the one whose mistakes cost the most rework.
Days 31 to 60. Install one review gate where shipping wrong hurts most.
Days 61 to 90. Keep one weekly sync ritual, about fifteen minutes, backed by a surface.

Handoff, Orchestrate, and Update arrive because the triangle demands them. Chartered roles need packets. A second gate forces you to map the flow. A ritual has nothing to discuss unless findings feed back.

Announce nothing. Install one thing.

Five ways this fails, and which skill repairs each

Multi-agent teams fail in a small, predictable set:

Failure	Tell	Repair
Shadow agents	Output no chartered role produced	Charter
Context rot	Plausible answers about a world that no longer exists	Handoff
Duplicate orchestration	Two workers, one task, no merge	Orchestrate
Review theatre	Near-100% approval; the reviewer cannot say what they checked	Review
Sync debt	Decisions live in chat; the surface is a lie	Sync

Diagnose your top two. Repair those. The rest shrink because the failures feed each other.

Where this sits next to SENSE and STACK

Different layer. Same seriousness.

AI Agents for Startup Strategy (SENSE) decides where to deploy agent systems and where the judgment boundary sits. Concept note: the SENSE introduction.
The Agentic Codebase (STACK) hardens the repository those agents live in. Concept note: the STACK introduction.
The Multi-Agent Organization (CHORUS) is the daily team protocol that runs the roster on a Tuesday.

SENSE without CHORUS deploys workers into a coordination vacuum. STACK without CHORUS versions the repo while the team still coordinates in hope. CHORUS without either still runs. The protocol fits in a README.

The named step lists live on the book pages and on the standalone CHORUS, STACK, and SENSE pages.

Who this book is for

the founder of a 2-to-15-person team with more agent roles than written contracts
the lead whose last customer-facing miss was unreviewed model output
the senior IC running a solo stack who keeps re-explaining the same brief
the operator who already chose what to deploy and now needs how the team runs

Skip it if you want a vendor tour or a prompt pack.

Buy the edition your team will actually use

Open the book page →

English editions (Amazon.com)

Format	Link
Kindle	Buy English Kindle
Paperback	Buy English paperback
Hardcover	Buy English hardcover

Dutch editions (Amazon.nl)

Format	Link
Kindle	Koop Nederlandse Kindle
Paperback	Koop Nederlandse paperback
Hardcover	Koop Nederlandse hardcover

Dutch catalogue page: boeken/multi-agent-organization.

Close the loop

You already run a multi-agent organization.

The chart just has not admitted it.

Charter before run. Handoff after charter. Review before ship.

Write the protocol. Then rent the tools.

The Multi-Agent Organization: Close the Coordination Gap With CHORUS