The Agent Team Model

The pattern: the coordinator at the center, the ten other specialists around it, and the clocks that fix the order in which they run; the ledgers beneath, and a human gate at every transition that changes a plan. The rung numbers mark which specialists may speak only at association.

The agent team model is the pattern by which a workforce planning function delegates a recurring loop of its own work to a set of specialized software agents: each agent does one kind of work, reads from and writes to a shared set of versioned files, is called in a fixed order by a coordinator on a clock, and hands its output to a human at every point where a plan would change. The model's central claim is that the quality of what such a team produces is set by the scaffolding around the agents, not by the capability of the model inside them, and that a general-purpose chat assistant, however capable, is not an agent team because it has none of that scaffolding. This page states the pattern's seven elements and the reasons for each, and gives the roster of specialists the rest of the series assumes.

Scaffolding over model

The applied machine-learning literature's most durable finding about production systems is that the modeling code is a small fraction of a deployment, and that the surrounding configuration, data handling, monitoring and process glue is where cost and hidden debt accumulate. The AI Scaffolding Framework states the same proportion for a workforce function as seven layers beneath the model, and AI Leverage Maturity in WFM Teams draws the planning consequence: architecture compounds while models depreciate. The agent team model is an instance of both. Everything on this page except the agents' reasoning is scaffolding, and all of it survives a change of model.

A practical corollary is that the team is specified in plain files: an identity and rules for each agent, a description of each ledger, the order of the clock, the gates. A planner can read the specification; an automation analyst can change it; the change is versioned. This is the discipline The Automation Analyst describes as the permission map, derived from the L2 step table's columns and nothing beyond them, and whose failure mode that page names as filling gaps from experience: every action an agent takes traces to a written step with an owner, and a gap in the table is returned to the process owner, not filled.

The seven elements

#	Element	What it is	Why it is there
1	The file system as context	Every planning object is a file in a known place: definitions, demand, forecast versions, supply, events, questions, plans, reports, and a change log. Agents receive no context except what those files hold	An agent that must be told the state of the world in a prompt is told it differently each time. A file is the same for every agent and every run, and a person can open it
2	Specialists, not a generalist	One agent per kind of work: reconcile data, score a forecast, match events, test a hypothesis, propose a reforecast, size a plan, review, report, exchange files	A specialist's rules fit on a page and can be checked. A generalist's cannot. Specialization also fixes who may speak at which rung of the causal ladder
3	A coordinator	An agent that owns no analysis: it runs the clocks, calls the specialists in order, holds the gates, and writes the run log	Order is a rule, not an emergent property; the coordinator is where the order lives and where a blocked run is visible
4	Clocks	Daily, weekly and monthly runs with fixed sequences, plus an intake door for anything that arrives by ask	The planning function already runs on these clocks. The agent team inherits them rather than inventing a cadence
5	Human gates	A named person approves anything that changes a plan: a reforecast, a capacity plan, a schedule publication, a severity, a release to a downstream system	Agents propose; people sign. The gate is enforced in the file format, not in policy: an unsigned proposal cannot be loaded
6	Graded numbers	Every figure carries [M] measured, [C] computed, [E] estimated or [A] asserted, and a computed figure inherits the weakest grade of its inputs	A carried assumption looks like a measurement in a spreadsheet. The grade is how the team is prevented from presenting one as the other
7	Living ledgers	Files are appended, never overwritten; a re-pull is a new version; every write is logged with who (agent or person) and why	The forecast that a decision was made against must survive the next refresh, for the reason Forecast Vintages gives

Elements 6 and 7 are developed on their own pages, Human Gates and Number Grades and Living Ledgers. The others are stated fully here.

The roster

The rest of the series assumes a roster of eleven specialists for the planning loop; the scheduling and real-time teams have their own rosters, on their pages. The names are generic and the roles are what matter; a function may merge two or split one.

Agent	Work	Rung it may speak at	Writes to
Coordinator	Clocks, order, gates, run log, backlog	—	Change log, run log
Librarian (intake)	Classifies every ask; pins definitions; opens register rows	—	Definitions, events, questions
Data engineer	Extract, reconcile, version; checks that sums match and definitions are cited	—	Demand, supply
Post-analyst	Forecast against actual by channel and interval; decomposes a miss into volume, handle time, mix and supply; control charts and regime detection	1, association	Forecast (variance), reports
Scout	Internal and external intelligence; proposes events with effect windows; screens for correlation	1, association	Events
Causal analyst	A causal diagram per question; confounder isolation; the test that would settle it; structural or transitional tag	2 and 3, intervention and counterfactual	Questions
Forecaster	Short- and mid-term reforecast; every assumption registered and versioned; scenario runs	—	Forecast
Capacity planner	Requirement hours to FTE to roster shape to gap; scenarios; the plan in the enterprise planning platform's shape	—	Plans
Evaluator	Adversarial review: no unlabeled carried assumption, grades present, rung respected, definitions cited; passes or blocks	—	Run log
Reporter	Answer-first daily note, weekly review, register report	—	Reports
Adapter (a family, one per boundary)	File exchange with the platform of record, the ACD, HR, finance and the planning platform; definitions and signature enforced at the boundary	—	Forecast, plans

The rung column is a rule, not a description. A post-analyst that writes "the go-live caused the handle-time rise" has broken it; the sentence it may write is "handle time rose from the go-live date, and the rise is out of control on the chart." The causal analyst may say more, after a diagram and a test, and the evaluator checks that it did.

Why a chat assistant is not an agent team

The distinction is not model capability. A single assistant given the same model can, in one long conversation, reconcile a file, score a forecast and draft a note, and on a good day the result reads well. What it lacks is structural.

• No fixed order. The assistant does the steps in the order the prompt suggests, which differs by prompt. The team's coordinator does them in the order the clock states, every run.
• No enforced rung. The assistant will answer "why" at whatever rung the question invites. The team's post-analyst cannot.
• No gate in the format. The assistant's output is text a person may or may not act on. The team's proposal is a file that a downstream adapter refuses to load without a signature.
• No inspectable trail. A wrong conclusion in a conversation is attributed to "the AI." In a team, the intermediate files say which specialist introduced the error, the property AI Leverage Maturity in WFM Teams calls inspectability.
• No durability. The conversation ends; the team's ledgers and specifications remain.

The general agents literature draws the line in the same place: an agent is distinguished from a program by autonomy, reactivity and a persistent environment it acts in, not by the sophistication of its reasoning. Recent evaluations of language-model agents make a narrower point that bears on planning: reported benchmark scores overstate reliability because they are measured on single attempts, and a task an agent completes on one run in two is not a task an operation can plan on. A team with a fixed order and an evaluator step is the form in which reliability can be measured per specialist, per run.

Worked example

On Wednesday 4 March 2026, the third day of the series example's migration, the daily clock runs the roster in order: the data engineer versions the day's actuals [M]; the post-analyst scores the previous day (voice handle time 452 seconds [M] against 412 seconds [C]) and writes the out-of-control chart as association; the scout proposes a go-live event at [A]; the forecaster proposes a reforecast whose handle-time assumption is now 452 seconds [M], with 412 seconds relabeled as carried [A]; the evaluator passes; the coordinator requests the gate; the planner approves; the reporter and adapter follow. The full run, step by step, is on The Short-Term Forecasting Loop with an Agent Team. Nothing in it required a person to type a number. Two things required a person to decide: whether the reforecast was acceptable, and whether the proposed event was real.

What would change this

The roster and the order are design choices drawn from a prototype, not findings; a function that measured a merged forecaster-and-planner specialist as equally reliable would be entitled to merge them. The claim that a chat assistant cannot substitute is structural and would be overturned only by an assistant that acquired the five properties above, at which point it would be an agent team.

How this connects

Living Ledgers is the data half of this pattern; Human Gates and Number Grades is the governance half. The Short-Term Forecasting Loop with an Agent Team shows the roster running one full daily clock, and Long-Term Planning Agents and the Plan of Record the monthly one. The pattern is the planning-function form of the general AI Scaffolding Framework, and the supervised-autonomy level of the ladder on AI Leverage Maturity in WFM Teams. The roles that build and check the team are on The Automation Analyst and The Agent Overseer.

Maturity Model Position

The pattern is a Level 4 construct that a function begins building at Level 3: the ledgers and definitions are Level 3 work, and the first specialist running on a clock is the Level 4 threshold. At Level 5 the gates would move from every proposal to exceptions only, on the evidence of catch rates the model is designed to produce.

See Also

• AI Agent Teams for Workforce Management — the series hub
• Living Ledgers — the file system as context, in detail
• Human Gates and Number Grades — gates and the two grade scales
• The Short-Term Forecasting Loop with an Agent Team — the daily clock end to end
• AI Scaffolding Framework — the general seven-layer form
• AI Leverage Maturity in WFM Teams — the leverage ladder and checkpoint placement
• The Automation Analyst — who converts a documented process into an agent
• The Ladder of Causation in WFM — the rungs the roster is bound by