Files
gcp_solution_architecture_a…/.agents/plans/iter-10-work-breakdown/steps/step-1-to-4.md
Jonathan Boniface b66d18fe94
Some checks failed
validation / verify (push) Failing after 10s
docs: record the first run of the work breakdown structure and of stated choices reaching Checks
Case 08: a plan of 9 work packages and 15 activities, 7.2 weeks against a
10-week window, with the customer's carrier circuit on the critical path. The
stated topology was built this time. The owner keeps the estimates as they
are: one experienced engineer, no contingency.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-11 21:04:28 +01:00

4.9 KiB

Steps 1 to 4 — Implementation Detail

Built on 2026-10-11. Not run with real models.

  1. A new Deliverable, work_breakdown, in the packaging Phase, after validation, produced by a new Specialist, the delivery planner (agents/delivery-planner.md, on the larger model). It is given the Target architecture and the deployment guidance in full.
  2. The planner writes the plan as data, work-breakdown.json: work packages (a hierarchy, each naming the Elements it delivers, its scope, what it produces and the skills it needs) and activities (work package, days, what each comes after, and whether the team or the customer owns it).
  3. Code checks the plan against the design (check_work_breakdown in app/workflows/work_breakdown.py). A submission is refused when an Element belongs to no work package, a work package names an Element that does not exist, an activity follows one that does not exist, activities depend on each other in a circle, a duration is not above zero or is over 120 days, or a work package holds neither an activity nor another work package.
  4. Code computes the schedule (schedule). An activity starts when everything it follows has finished. The team's activities also wait for a free person: no more run at once than the team has people, with the longest remaining chain placed first. What the customer owns takes nobody from the team. The critical path is every activity with no slack.
  5. Code writes the three views into the Deliverable when it is submitted, as it writes the architecture's tables:
    • the WBS dictionary, each work package with the Checks on its Elements as what makes it done, read from the Target architecture;
    • the activity sequence, with each activity's start and finish day and whether it is critical;
    • the timeline, a chart in text along a relative axis from Day 0, the team's work and the customer's drawn differently and the critical path marked.
  6. The delivery terms come from the agreed scope: team size, working week, delivery window and start date. A team size the scope does not give is 2, and the Deliverable says it was assumed.
  7. The plan is checked against the delivery window. Over it, the Deliverable says by how much and names the critical path. Nothing shortens an estimate to fit.
  8. What the customer owns is listed with its lead time and the assumption behind it.
  9. With a start date in the scope the activity sequence also gives calendar dates, skipping weekends, and a Mermaid Gantt chart is attached as timeline.mmd. Without one there are no dates to draw, and the text chart is the timeline.
  10. The guide carries the timeline. The guide and the plan are produced side by side, so neither reads the other; when an Execution's files are written, the plan's timeline is added to the guide as "Delivery plan". execution.json records the plan's length, team, whether it fits the window and its critical path, and the packaging Approval says the same.

Differs from the spec:

  • Milestones for fixed-price work are not built.
  • An assumption in the scope is not linked to an activity by id. An activity the customer owns carries its own assumption as text.
  • The timeline chart is text. A Mermaid Gantt chart needs dates, so one is drawn only when the scope gives a start date.
  • The guide does not depend on the plan. Making it wait would have put the plan on the path of every Execution's last step; adding the timeline in code keeps them independent.

Durations are a model's estimates. The Deliverable says so.

Tests: tests/execution/test_work_breakdown.py (16).

First run with real models, 2026-10-11

case-08-landing-zone, results 20261011T185421Z. The delivery planner made one model request, about $0.32.

  • The plan: 9 work packages under one parent and 15 activities. Every one of the design's Elements is in a work package, and each work package lists the Checks that make it done.
  • The schedule: 36 working days, 7.2 weeks, with the team of 3 the scope states. Inside the 10-week window with 2.8 weeks to spare.
  • The critical path is the customer's: ordering the interconnect circuits, 30 days, then the routers and attachments, then validation in the three environments the scope names. The team's own work is done by about Day 13 and waits.
  • The timeline was added to the solution guide and its numbers are in execution.json.

To look at:

  • The estimates look short for an enterprise landing zone. Two days for the hub and spoke networks, one for the pipeline and state, two for the resource hierarchy. The planner was told to estimate for one engineer who knows the provider and not to pad, and it did. Whether to ask for a contingency, or a range, is the owner's call.
  • The plan's length rests on one assumed number, the 30-day carrier lead time, which the planner chose. It is shown as the customer's and as an assumption.
  • No start date was in the spec, so there are no calendar dates and no Mermaid chart.