Known work
When the agent expands work you already know
Knowing the work helps you spot errors. It does not make constant supervision free.
Written and researched by Trần Thiện Học · Independent practice-led research on human and AI work
Human situation
A one-file fix becomes a repository refactor. A caption becomes brand strategy. Work you understand keeps adding scope.
This article reflects practice-led research. Observations, measurements and open questions are labeled separately.
What usually goes wrong
Useful exploration quietly replaces the finish line. The agent may be correct at each step; the work is simply no longer the work you meant to finish.
Why the work feels heavy
You must remember the original outcome, compare each addition against it and decide whether the extra branch earns its cost.
What Dwi changes
Conduct names the outcome. Lean selects the smallest sufficient path. Budget sets a checkpoint before expansion. Arc adds lanes only when they are truly separate. Evidence returns validation and unresolved risk.
What Dwi cannot solve
Dwi cannot determine the correct product decision without domain judgment. It preserves the place where that judgment is applied.
A clearer scope does not decide which product trade-off is worth making.
A possible workflow
Name one finish line, one mutable scope and one evidence request. Bring the agent back to the checkpoint before a new branch becomes a new project.
Evidence to ask for
Request the changed artifact, the checks that ran, the scope that was not touched and the risks that remain.
- Changed artifact
- Checks and diff
- Untouched scope
- Remaining risk
A safe next step
Write the finish line in one sentence, then let the agent work until that sentence is true.
Dwi by thienhoc
Return to the finish line.
Run one scoped change, inspect what moved and stop before useful drift becomes a second project.