Question
Why can a team with more work underway feel busier yet deliver later?
Background
Delivery work is shaped by capacity, dependencies, blockers, milestones, and risk. Agent-assisted work adds a twist: it may speed implementation while placing more demand on the next shared step—review, integration, product clarification, test, or release.
Two queueing facts
L = λW: average active items equal throughput times average time in the system. With stable throughput, more work in progress means longer lead time.
Queueing models show waiting time rises sharply as utilization approaches full capacity. Variation in arrivals and service time amplifies that effect.
What this changes
The goal is not to maximize every person's utilization. It is to protect the actual constraint from starvation while limiting work that is only waiting for it. An idle-looking gap can be reserve capacity for an urgent unblock, an unexpectedly difficult review, or a necessary decision.
Illustrative arithmetic
If a workflow completes about five items a week and has ten items active on average, Little's Law corresponds to an average two-week time in the system. Five active items at the same throughput corresponds to one week. This describes a relationship, not a promise that reducing work in progress alone causes the result.
Codex inference
When agent assistance makes implementation faster, inspect whether the bottleneck has migrated. If a reviewer, decision-maker, or test environment is now the constrained step, generating more work can enlarge that queue without improving end-to-end delivery.
A useful experiment
- Choose one workflow and define its start and completion points.
- Record daily active-item count and each item's time in implementation, review, clarification, test, and release.
- Compare lead time with concurrent work, and find where time accumulates.
- Use the result to test a smaller work-in-progress limit at the suspected constraint.
Keep the conclusion local. Scope changes, batching, and poor status data can all make the simple model misleading.