Dwi

Bounded lanes

Place stronger reasoning where it pays.

Arc is model-agnostic. A strong reasoning model can shape the plan and integrate bounded execution cells.

Written and researched by Trần Thiện Học · Independent practice-led research on human and AI work

Human situation

Arc is model-agnostic. A strong reasoning model can shape the plan and integrate bounded execution cells.

This article reflects practice-led research. Observations, measurements and open questions are labeled separately.

What usually goes wrong

Parallelism is added before the dependency graph is understood, so the root spends its time repairing overlap.

  • Strong model at Root
  • Lower-cost bounded cells
  • Independent review
  • Human final authority

Why the work feels heavy

The person must coordinate model choice, lane boundaries, merges and proof while the apparent speed keeps moving.

What Dwi changes

Root can handle decomposition, boundary definition, conflict resolution and integration. Lower-cost models can handle narrow implementation, bounded research, data transformation and repeatable checks.

What Dwi cannot solve

Savings appear only when cell scopes are clear, retries controlled and integration does not erase the savings.

A stronger model mix cannot remove dependencies that were never separated.

A possible workflow

Root marks dependencies first. Arc assigns only disjoint work, receives lane artifacts and keeps the merge decision in one accountable place.

Evidence to ask for

Request each lane's artifact, dependency assumptions, merge conflicts, retry count and the reason a sequential baseline was not enough.

  • Lane artifact
  • Dependency map
  • Merge conflicts
  • Human gate

A safe next step

Compare one bounded lane with the sequential baseline before adding a second lane.

Dwi by thienhoc

Compare the lanes before adding another.

Keep parallel work only when the dependency shape and the returned evidence make the gain legible.