A Roadmap Pattern for Agent Teams in WFM

A roadmap pattern for agent teams in WFM is a generic sequence for bringing agent teams into a workforce planning function's own work: the short-term forecasting loop first, with planners approving; the question register second; integration adapters third; the monthly and annual plan fourth; scheduling and real-time last. The page's claim is that the order follows from where failure is contained and where standardization already exists, not from where the savings look largest, and that the one measurement a function must make before it plans on an agent team, what a planner-hour currently covers, is the measurement almost no function has. This page gives the sequence, the reason for each step, the planner-hour measurement, and what the pattern deliberately leaves out.

The sequence

Step	What is handed over	Why here	Exit evidence
1	The short-term loop, with the planner at the gate	Failure is contained; the loop is already half-documented; the planner who approves is the person who can correct the specification; the daily clock produces a variance record every day, which is the evidence every later step reads	Thirty consecutive daily runs with the planner gate exercised; the readiness check on Agent Team Readiness for a Planning Process passed; a first catch rate published per The Agent Overseer, with its injection rate and sample size beside it
2	The question register and the intake door	The loop's unexplained misses need somewhere to go; the register can be run by hand before the causal analyst is built; it is the first place the function's knowledge starts to accumulate rather than repeat	Twenty rows closed with graded cards; one card reopened by its own "would change this" observation; the first data pull refused for want of a decision
3	Integration adapters, file first	The loop's forecast must reach the platform of record and the ACD's actuals must reach the demand ledger without a person copying files; the signature rule becomes mechanical here	Every crossing logged; one batch window in production passed with the gate unapproved and nothing written, or one unsigned version refused; the mapping file's rejection log opened
4	The monthly and annual plan	The stakes are higher and the cycle is slower, so the loop's evidence should exist first; the plan reads the register's structural tags and the forecast ledger's versions, which steps 1 and 2 produce	One signed plan of record exported from a scenario pack; one move declined with its price recorded
5	Scheduling agents	Publication touches every worker; the checker needs a signed plan to reconcile against, which step 4 produces	A check record that caught a stale forecast version; a bid pack with its register
6	Real-time agents	The action gate needs a real-time analyst still practiced at the day, and the events loop needs a scout to receive its proposals; it is last because it is where an error reaches a customer fastest	An incident record with evidence attached; an event confirmed from a real-time proposal and matched by the daily loop

The sequence has a shape: it starts where the function's work is most its own and moves outward to where the work touches workers and customers. That is the opposite of the order most automation business cases propose, which start where the hours are largest. The hours are largest in scheduling and real-time; the evidence is earned in forecasting.

Why not start where the savings are

Two findings from the field evidence on generative assistance bear on the order. The first is that gains are largest for the least experienced workers and on work inside the model's competence, and that the boundary of that competence is jagged and not obvious in advance. A planning function does not know in advance which of its processes fall inside the boundary; the short-term loop is where it finds out at the lowest cost. The second is that measured agent reliability on multi-step tasks falls with task length, and that the length agents complete reliably has been rising on a measured trend rather than arriving all at once. A daily loop of ten short steps with a gate is the task shape that evidence favors; an unattended real-time layer is not, yet.

The human-factors argument is the same one every page in the series returns to: the person left at the gate must still be practiced at the work, and the order above keeps the planner practiced on the forecast, the scheduler on the schedule and the real-time analyst on the day, for as long as each gate stands.

Measuring what a planner-hour covers

Before a function plans on an agent team, it needs one number it almost never has: what a planner-hour currently produces, on which process, at which grade. Without it, every claim that the team "saves" time is an assertion, and every plan built on that claim is [A].

The measurement is small. For one process, the short-term loop, each planner logs for two weeks: minutes per day on each step of the L2 table; which steps were skipped when time ran short; which figures in the output were carried rather than measured. Three results follow. The coverage of the process: the share of its steps actually run on a typical day, which is usually below what the function believed. The grade profile of the output: the share of figures that are [M] rather than [A]. And the hours, by step. The agent team's first claim is then not "it saves N hours" but "it runs every step every day and the output's grade profile moves from this to that," which is a claim the ledgers can score. The hours claim comes after, from the same log repeated with the team running.

The measurement follows the ordinary discipline: define the decision (whether to plan on the team), state what is uncertain (what the hour covers), and measure that. A function that skips it will find its agent team credited with hours it never freed and blamed for a coverage gap it inherited.

What the pattern leaves out

• A headcount target. The pattern produces evidence about what the team covers; what the function does with freed planner hours is a placement decision made on that evidence, and The Hardening Residual describes why the remaining work per planner gets harder as the routine leaves.
• Unattended execution. No step in the sequence lifts a gate. Lifting one is a change to a file format on the evidence of a published catch rate, per Human Gates and Number Grades and The Agent Overseer.
• Customer-facing agents. The sequence is for the planning function's own work. Agents that handle contacts are placed by The Agentic Handover Gate and planned as supply by Agentic AI Workforce Planning, on their own roadmap.
• Dates. The steps are ordered, not scheduled. A function's dates depend on where it enters, which The Agentic Journey Map says is wherever its definitions and documentation have reached.

Worked example

The series example's function, entering in February 2026 with a migration as its clock. Step 1 runs from Monday 2 March; the planner-hour log had run for two weeks in February and showed the reforecast's L2 steps run in full on 6 of 10 days, with the handle-time assumption carried on every one of them. By Friday 3 April the loop has run 25 consecutive business days, the gate has been exercised on each, a first catch rate has been published per The Agent Overseer (9 of 10 known errors detected, from 6 injected and 4 found by deep verification of a 5 percent sample [C from the overseer's log]), and the grade profile of the daily note has moved from 40 percent [M] to 85 percent [M] [C from the log]. Step 2's register opened on 23 February with Q-001, before step 1 began, as remediation of readiness question 4 on Agent Team Readiness for a Planning Process; the roadmap's step numbers order handovers, not artifacts. By 10 April it holds eleven closed rows; Q-011, the training-pull card, is the one the librarian returns in one line when a manager asks the same question on 10 April. Step 3's forecast adapter went live in week one as a file exchange and wrote nothing on 5 March when the gate had not been approved by the batch window. Step 4's March cycle signed the phase 2 plan on 27 March with one move declined and priced. Steps 5 and 6 begin in April, the checker on the week-of-30-March schedule and the real-time issuer on 8 April, before step 1 had reached the thirty runs its exit evidence asks for; the function accepted that the checker's first weeks would run beside a loop whose catch rate was one period old, and recorded the acceptance. The sequence does not require steps 5 and 6 to wait for step 4 to be perfect, only for its plan of record to exist.

None of the dates is a recommendation; they are the example's, and they exist so that the pages agree.

What would change this

The order is a design argument from containment and evidence, and it is testable. A function that started at scheduling, with a checker and a publication gate, and reached a published catch rate faster than one that started at the forecast, would weaken the argument for step 1 first. Evidence that the planner-hour log was not worth its two weeks, because functions turned out to know their coverage already, would drop the measurement section; the page expects the opposite.

How this connects

The roadmap sequences the series: step 1 is The Short-Term Forecasting Loop with an Agent Team, step 2 is The Question Register and Knowledge Base, step 3 is Integration Agents, step 4 is Long-Term Planning Agents and the Plan of Record, steps 5 and 6 are Scheduling Agents and Real-Time Agents; each step's entry condition is the check on Agent Team Readiness for a Planning Process. The wider route a function follows to reach agentic execution at all is The Agentic Journey Map, and the technology path beside it is Technology Journey from Level 2 to Level 5.

A function that adopts this sequence as a chartered program, with one owner as a seat, a first ninety days and the planner-hour measurement above as the instrument its founding claim is graded against, is described on AI Agent Program for a Resource Optimization Center; the rungs a team on one book climbs while the sequence runs, and the exit evidence each step's row above becomes on a ladder card, are on The Agent Team Ladder: Alpha to Production.

Maturity Model Position

Steps 1 to 3 move a planning function from Level 3 to Level 4 on the WFM Labs Maturity Model™: a daily loop with versions and grades, a register, one set of definitions enforced at the boundary. Steps 4 to 6 are Level 4 practice across the lifecycle. The pattern stops before Level 5 by design; the evidence its gates produce is what a function would use to decide whether to go there.

See Also

• AI Agent Teams for Workforce Management — the series hub
• Agent Team Readiness for a Planning Process — the entry check for every step
• The Short-Term Forecasting Loop with an Agent Team — step 1
• The Question Register and Knowledge Base — step 2
• Integration Agents — step 3
• The Agentic Journey Map — the wider route this sequence runs inside
• The Hardening Residual — why remaining work per planner gets harder
• Technology Journey from Level 2 to Level 5 — the technology path beside the sequence
• AI Agent Program for a Resource Optimization Center — the sequence chartered as a program, judged on detection and catch rate
• The Agent Team Ladder: Alpha to Production — the exit evidence per step recorded as rungs on a ladder card