ROSTER is for one person running several AI agents. CHORUS is for a small team of people who share work with AI agents. An AI agent here means software that takes a goal, chooses its own steps and uses tools such as search, email or files. ROSTER comes from The Agent Dream Team and turns each agent into a specialist with a job, a target and a weekly review. CHORUS comes from The Multi-Agent Organization and makes work, context and review travel cleanly between people and agents in a team of 2 to 15. Both books are by Len P. van der Hof and available now. You do not need either one to use this page.
The quickest way to choose: count the people who can say no to a piece of work before it goes out. If the only name is yours, start with ROSTER. If there are two or more, you need CHORUS as well. The Agent Dream Team puts it in one line: “CHORUS scales a team. ROSTER scales a person.”
What does each framework ask you to do?
ROSTER: six decisions for every agent
- Roles. One job per agent, written in a one-page charter (a job description) that says what the role owns and what it does not.
- Objectives. A result specific enough to grade on a Friday without argument.
- Skills and tools. Only the access that job needs, and nothing else.
- Triggers. The event that starts the agent, decided once.
- Evaluation. A written scorecard, graded weekly, with you as the final check.
- Rotation. Upgrade, merge or retire a role when the scorecard or the work says so.
The ROSTER framework page has the book’s one-line version of each step.
CHORUS: six skills for work that crosses between workers
- Charter. A seven-field contract for every worker, human or agent: purpose, inputs, outputs, authority ceiling (what it may never decide), success criteria, review requirement, escalation.
- Handoff. A six-field packet that travels with each task: goal, constraints, inputs, provenance (where each input came from), acceptance criteria, deadline.
- Orchestrate. Steps in order, work split across parallel agents, results merged by a named owner, and a named person to escalate to.
- Review. A gate where output passes only when it meets written criteria, not because it sounded right.
- Update. What reviews find flows back into prompts, tools, procedures and charters, so the next run improves.
- Sync. One place that shows who owns what and what is blocked, read in a short weekly meeting.
What is CHORUS? fills in a charter and a packet for the same Friday task.
Where do they differ?
The difference is not the letters. It is who holds the accountability. This table follows the comparison in The Agent Dream Team, in plain words.
| Question | ROSTER (one founder) | CHORUS (team of 2 to 15) |
|---|---|---|
| Who is accountable? | You alone | Several people, working with agents |
| Who reviews the output? | You are the only human reviewer | Team members review each other’s work |
| Who can overrule an agent? | Only you | Any accountable person; people can also overrule each other |
| Where do handoffs go? | From agent to agent, and back to you | Between people, between agents, and between the two |
| How does everyone stay in step? | A weekly routine you run alone | Team rituals and a shared status board |
| Who catches a failure? | You, in your Friday review | The team spots it and reroutes the work |
The book is explicit that a solo founder running ROSTER is not running a smaller version of a CHORUS team. One person is manager, reviewer and final gate at once. That is a design fact to plan around, not a flaw. It is also why the weekly review carries so much weight in ROSTER: nobody else will catch the drift.
Where do they overlap?
The book says the two frameworks share their DNA. Both charter every worker, both insist on written handoffs, and both require a human review before output ships. The details differ.
| Part | ROSTER | CHORUS |
|---|---|---|
| Charter | Six fields: role name, mandate (owns and does not own), input, output, permitted tools, escalation rule | Seven fields: purpose, inputs, outputs, authority ceiling, success criteria, review requirement, escalation |
| Handoff packet | Four fields (task summary, relevant context, open questions, output files), labelled with sender, receiver, project and version | Six fields: goal, constraints, inputs, provenance, acceptance criteria, deadline |
| Review | Your weekly scorecard pass | A gate with a named person who can fail the work |
So “which framework has handoffs?” is the wrong question. Both do. The real question is who stands on the other side of the handoff, and whether anyone besides you can refuse it. For the full seven-field contract with a worked example, see What is an AI agent charter?
A worked example: one consultant, then a team
Imagine Maya, a hypothetical consultant who writes client proposals.
Stage one: Maya alone. She runs two agents. A research agent gathers sources about the client; a draft agent writes the proposal. One week the research notes contain a pricing recommendation nobody asked for, and the draft agent copies it into the proposal. That is a role problem: the research agent’s charter never said “does not advise on price”. The ROSTER fix is one line under Roles and one check under Evaluation: no pricing advice in research notes.
Stage two: Maya with a team. She hires a writer and an account manager. A month later the writer produces a strong proposal for a scope the client dropped two weeks earlier. The research agent did its job; the writer worked from an old brief. That is a handoff problem, and it now runs between people. The CHORUS fix is a packet with provenance (which brief, which version, supplied by whom) and a named person who must accept it before writing starts. The account manager owns that acceptance, and the shared board shows which brief is current.
Notice what did not change. ROSTER’s packet carries a version label, so it could have caught the stale brief when Maya worked alone. The team did not create the need for a handoff. It multiplied the people who could own one, which is why someone now has to be named. That matches the book’s own advice: founders who grow past a company of one typically need both frameworks, ROSTER for their own agents and CHORUS for the people they bring on.
When something breaks, which one do you fix?
Headcount tells you which framework to start with. The failure tells you which part to repair first.
| What you see | What is missing | Where the fix lives |
|---|---|---|
| An agent does work nobody asked for | A written limit on its role | ROSTER: Roles (CHORUS: Charter) |
| You rewrite the output every week | A result you can grade | ROSTER: Objectives and Evaluation |
| An agent can reach systems its job never needs | A tool boundary | ROSTER: Skills and tools (CHORUS: the authority ceiling in the Charter) |
| The next worker starts from an old version | A handoff with version and source | ROSTER’s handoff packet, or CHORUS: Handoff |
| Two drafts arrive and nobody owns the merge | A named owner for the merge | CHORUS: Orchestrate |
| Work ships because nobody said no | A gate with a named person who can fail it | CHORUS: Review (alone: your Friday scorecard) |
| Nobody knows what is blocked | One shared status view | CHORUS: Sync |
| An agent still does last quarter’s job | A decision to change or retire it | ROSTER: Rotation (CHORUS: Update) |
If most of your rows land in the bottom half, the problem is coordination rather than any single agent. Why multi-agent systems fail walks through those breaks in order.
Common mix-ups
“CHORUS is ROSTER for bigger companies.” No. The Agent Dream Team says it outright: CHORUS is not ROSTER at scale. ROSTER assumes one human coach; CHORUS assumes several accountable people.
“CHORUS is for enterprises.” Also no. The Multi-Agent Organization is scoped to teams of 2 to 15 people. Large-organisation design is outside the book.
“A solo founder needs no handoffs.” Work still moves from one agent to the next, and ROSTER has its own packet for exactly that.
“Either one is a software product.” Neither is. A tool can run tasks in sequence without carrying the right context or giving anyone the authority to reject the result. A plain shared document can hold a good charter or packet.
Try this today: the 15-minute “who can say no” check
Pick one piece of work that went out last week: a proposal, a report, a set of release notes.
- Write down every worker that touched it, human or agent, in order.
- For each step, fill in three columns: who sent it, who received it, who could have rejected it.
- Read the last column.
- Only your name: you are in ROSTER territory. Write the “does not own” line for the agent that surprised you most.
- Several names: you need CHORUS. Write a six-field packet for the handoff that failed or felt shaky.
- A blank: nobody could reject the work. Name that person before the next run.
Cite this:ROSTER vs CHORUS: AI agents for one person, or for a team?Len P. van der Hof. https://lenvanderhof.com/en/blog/roster-vs-chorus/ ·