What is ROSTER? How one person manages a team of AI agents

ROSTER is a management framework for a one-person business that runs several AI agents. An AI agent here means software that takes a goal, chooses its own steps and uses tools such as search, email or a shared folder. ROSTER turns “I use a few AI tools” into “I run a small team of named specialists” by making six decisions for every agent: Roles, Objectives, Skills and tools, Triggers, Evaluation, Rotation.

The framework comes from The Agent Dream Team by Len P. van der Hof, available now in English and Dutch. You do not need the book to use it. This page gives you the six decisions in plain words, one role filled in from top to bottom, the book’s rules on how many agents to run, and a fifteen-minute audit you can do today.

What problem does ROSTER solve?

Most solo founders start with one chat window that does everything: research on Monday, a client email on Tuesday, the invoice list on Wednesday. The book names four ways that single thread degrades.

• Task-type collision. Research, writing and admin share one conversation, so the habits of one job bleed into the next.
• Memory pollution. Old material (last month’s client, a rejected draft) piles up and drowns out what matters for today’s task.
• Tool sprawl. One assistant holds every permission, so a research question sits one step away from a tool that can send or delete.
• Quality variance. There is no written standard, so you judge each output by mood and cannot tell good work from bad.

ROSTER’s answer is the one small companies already use: split the work into jobs, and write down what each job is.

The six decisions, in plain words

1. Roles. Give each agent one job and a one-page charter (a written job description). The book’s charter has six fields: role name, mandate (what the role owns and, explicitly, what it does not), input (what it receives), output (what it delivers), permitted tools, and an escalation rule (what it must hand back to you instead of attempting).
2. Objectives. Make the role measurable. “Deliver three sourced competitor profiles by Thursday” is an objective. “Help me with research” is a wish. A role without an objective is a job title.
3. Skills and tools. Give the role only the access its job needs, and cut the rest. This is least privilege, the security rule that every account gets the minimum rights its job requires. Enforce it in the tool settings; a sentence in the prompt removes nothing.
4. Triggers. Decide which event starts which agent: a new email from a prospect, a draft landing in a shared folder, a list you approve. You decide once, instead of choosing a chat window every time work arrives.
5. Evaluation. Grade the output against a written scorecard of three to five criteria, every week, with you as the final check before anything goes out.
6. Rotation. Roles are not permanent. When a role keeps failing its scorecard, or the work it was built for changes, you upgrade it, merge it with another role, or retire it.

The six form a loop, not a form you fill in once. The book’s own examples: a trigger that wakes the wrong agent is a role-design problem, and a scorecard that cannot catch bad output is an objectives problem.

What does one role look like, filled in?

Here is a hypothetical role for a one-person software company: an agent that drafts release notes (the short “what is new” text customers read after an update). All of it fits on one page, which is the book’s test for a charter. If it does not fit on one page, the mandate is too broad.

Decision	What you write down
Roles	Release-note drafter. Owns: turning the approved release list into customer-facing notes. Does not own: deciding what shipped, publishing, promising dates. Hands back: any item it cannot match to the list.
Objectives	Each Tuesday, a draft in which every change matches the approved list, nothing unreleased appears, and gaps are flagged instead of filled in.
Skills and tools	Reads the approved list and the product docs. Writes to one drafts folder. No publishing login, no email.
Triggers	You mark the release list “approved”. No approval, no run. A weekly calendar slot does not prove there is anything to announce.
Evaluation	Three checks each week: every claim traces to the list; no unreleased feature; unclear items flagged, not guessed. Once a month, plant one unreleased item in the input and confirm the draft leaves it out.
Rotation	One failed week: read the last three drafts and find the cause. Two failed weeks in a row: rewrite the prompt from the charter. Still failing two weeks later: retire the role and write the notes yourself.

Look at the “does not own” line. The book calls the explicit list of what a role does not own the line founders skip. It is also the line that stops a drafting agent from announcing a feature you have not shipped.

For a second filled-in charter, see the AI agent charter guide. It uses the seven-field version from the sister book on team coordination, which overlaps with the six fields above.

How many agents should you run?

Fewer than you think. The book’s starting point is three roles, which it calls the Minimum Viable Roster:

• a Research Specialist that gathers, summarises and cites, and never drafts or advises;
• a Draft Editor that writes in your house voice, and never researches or schedules;
• an Ops Coordinator that schedules, tracks and logs, and never writes client-facing text.

Three to five roles is the working range for a founder who reviews everything personally. Around eight, the book says, the founder stops being a coach and becomes the bottleneck again: every role you add is a role you have to read.

Six decisions do not mean six agents. Several roles can run on the same AI model with different instructions, access and scorecards.

There is also a bill. In its own write-up of a research system, Anthropic reported that a multi-agent setup beat a single agent by 90.2% on the company’s internal research evaluation, and that multi-agent systems use about 15 times as many tokens (the units AI providers bill by) as a normal chat. That is one vendor measuring its own system, not proof that more agents beat fewer. It does show that a roster is a spending decision as much as a design decision.

When does a role get fixed, merged or retired?

The book gives rules you can write on the charter before anything goes wrong.

• One week below your scorecard floor (the lowest score you will accept) starts a diagnosis, not a verdict. Read the last three outputs and look for the cause.
• Two weeks in a row below the floor, or four of the last six, start a decision: upgrade the role (rewrite the prompt, try another model, or add a missing tool), merge it with a role doing overlapping work, or retire it.
• Before you blame the role, check the model underneath it. If the provider changed or retired it, treat that as a forced move: switch, re-test, and only then judge the design.
• Once a quarter, review the roster as a whole: which roles are still needed, which overlap, which work you still do yourself, and what you would build from scratch today.

The author is plain that these counts are his operating rules, not research findings, and the book states that no study measures whether rotating agent roles improves a solo founder’s output. Pick your own thresholds. The point is to write them down in a calm week, so that in a bad week you read a rule instead of negotiating with yourself.

What ROSTER is not

Not software. It is a set of decisions that works with any AI tool. A charter can live in a text file; a system prompt is only one way to encode it. The step limits and logs in Using AI agents effectively are how you enforce the charter in a real tool.

Not CHORUS. CHORUS (Charter, Handoff, Orchestrate, Review, Update, Sync) is the sister framework from The Multi-Agent Organization, also available, written for teams of 2 to 15 people who share work with agents. The Agent Dream Team puts the split in one line: “CHORUS scales a team. ROSTER scales a person.”

Not a productivity guarantee. Named roles make failures visible and fixable. They do not prove the agents will do better work.

How does ROSTER run week to week?

The book pairs the six decisions with a thirty-minute weekly routine: a Monday charter review, a midweek check that the triggers fired as intended, and a Friday scorecard pass in which you grade each role. In a company of one, that Friday pass is the main systematic review in the whole setup. Skip it, and quality can drift until a deliverable fails in front of a client.

Try this today: a 15-minute ROSTER gap audit

This is a shortened version of the book’s first exercise.

1. List every AI workflow you ran in the last two weeks: research, writing, client email, scheduling, review.
2. Give each workflow a candidate role name. Group workflows that belong to the same specialist.
3. Score each role against the six decisions: Y (in place), P (partly), N (missing).

Role (hypothetical)	Roles	Objectives	Skills and tools	Triggers	Evaluation	Rotation
Client research	P	N	N	Y	N	N
Newsletter draft	Y	P	N	P	N	N
Invoice reminders	N	N	P	Y	N	N

4. Count the Ns. Each one is a gap in how you run your agents.
5. For the role with the most Ns, write two lines before you close the tab: what it does not own, and one objective you could grade on Friday.

The ROSTER framework page keeps the six steps on one screen if you want them next to your list.

Cite this:What is ROSTER? How one person manages a team of AI agents.Len P. van der Hof. https://lenvanderhof.com/en/blog/what-is-roster/ · Published 26 September 2026.