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.