Schedule generation

Proposal · not implemented

The proposed flow: the manager describes the store once, the agent drafts a schedule for the period, and the manager reviews, adjusts and publishes it.

The manager’s path

1. Describe the store. Opening hours, roles (for example cashier, key holder, floor), and the rules that apply: minimum rest, maximum hours, breaks. This is done once and updated when something changes.
2. Add the team. Employees, their roles, contracted hours and availability. See Input data for the fields we expect to need.
3. Provide traffic. How busy the store is by hour and day, from footfall counters, till receipts, or the manager’s own estimate.
4. Ask for a schedule. The agent drafts a schedule for the chosen period, for example the next two weeks.
5. Review. The draft shows coverage against traffic hour by hour, highlights gaps and overtime, and explains its choices in plain language (“Ben starts at 10:00 because the lunch peak needs three people”).
6. Adjust and publish. The manager can edit shifts directly or ask the agent in words (“give Chloe fewer closing shifts”). Publishing sends the schedule to the team.

Success looks like: a published schedule where every busy hour is covered, no rule is broken, and the manager spent minutes reviewing rather than hours building.

When the agent cannot find a valid schedule

Sometimes the rules cannot all be met: too few people for the traffic, or availability that leaves a gap. The agent should not hide this. It should say which hours are uncovered, which rule blocks them, and what would fix it (“one more person on Saturday afternoon, or allow Anna 2 extra hours”). The manager decides.

Agent or solver?

Assigning people to shifts under hard rules is a classic optimisation problem, and dedicated solvers are good at it. Our working assumption is that a language-model agent should not do the arithmetic itself. It should understand the manager’s and employees’ requests, turn them into constraints, call a solver, and explain the result. This is a hypothesis to test, recorded in Open questions. What the solver is, what the agent would ask of it and whether it needs MCP is on Agent and solver.

After publication

Changes after the schedule is out go through the Employee assistant or through the manager directly. Every change is re-checked against the same rules as the original schedule.