Sub-agents took the operational work out of workplan development.

The problem.

Developing a workplan sounds like planning. In practice it was assembly: pulling asset history from one system, crew availability from another, outage windows from a third; drafting the work package; checking dependencies against everything already scheduled; compiling permits and switching references; reconciling the schedule every time any upstream input moved.

Planners with decades of field knowledge were spending the bulk of their week on copy-and-check work a system should have been doing — and the cycle time from request to released package stretched across weeks when inputs churned.

What we built.

Not one big agent — a set of small ones, each with a single job a planner could verify at a glance:

• A package drafter that assembled the work package skeleton from the systems of record — asset history, job templates, prior similar work — with every field traceable to its source.
• A dependency checker that flagged conflicts with scheduled outages, crew assignments, and material lead times before a human ever reviewed the draft.
• A paperwork compiler that gathered the permit, clearance, and reference documents each package type required, and listed what was missing rather than guessing.
• A reconciliation agent that watched upstream changes and proposed schedule updates — proposed, never applied; a planner accepted or rejected each one.

How it was governed.

Every agent drafted; no agent released. Planners stayed the named approvers on every package, and each agent's output carried its sources, so checking a draft took minutes instead of rebuilding it. What the agents couldn't establish, they said so — a missing document was a listed gap, not an inference.

The impact.

The team's own measure of success was quieter: rework fell, because dependency conflicts surfaced before release instead of in the field, and the reconciliation queue meant schedule changes stopped arriving as surprises.

Outcomes from the founding team’s engagement are rounded and anonymized. Time-study methodology, package-type scope, and baselines are available in a review conversation.