The decision workbench
Put the tradeoff
on the table.
Three local tools for planning an AI-assisted change. No uploads, account connection, vendor scoring or automated release decisions.
01 / Planning estimate
Count accepted changes.
Count the review.
Compare two hypothetical batches of the same task type. Include time for rejected attempts. A change counts as accepted only after your review criteria are met. The defaults are illustrative inputs, not measured outcomes.
Formula, assumptions and what this excludes
Baseline cost = tasks × (implementation + review) ÷ 60 × hourly cost. Assisted cost = tasks × (implementation + review + rework) ÷ 60 × hourly cost + tool cost. Cost per accepted change = batch cost ÷ accepted changes. Net saving = baseline batch cost − assisted batch cost.
Both batches use the same attempted-task count and labor rate. Baseline implementation must include baseline rework; AI rework is separate. Include total human-attended minutes, not background machine time, in labor inputs. Tool cost can include subscriptions allocated to the batch, model usage and execution fees. Currency is user-defined; no exchange rates or current vendor prices are supplied.
Task mix, variance, long-term maintenance, incidents, opportunity cost and quality beyond your acceptance rule are excluded. Different accepted counts make total batch savings misleading; compare cost per accepted change and investigate the rejected work. This is planning arithmetic, not evidence of causal productivity gains.
02 / Qualitative comparison
Where does the
work belong?
Choose the strongest constraint. This compares workflow classes, not named vendors. Recommendations are editorial starting points; configured permissions and deployment details can change the answer.
| Surface | Useful when | Inspect first | Handoff evidence |
|---|---|---|---|
| Editor | You steer and review while working. | Extension, context and command access. | Diff + focused tests + manual checks. |
| Terminal | Commands and repository artifacts define the task. | Shell authority, working directory and secrets. | Reproducible commands + outputs + patch. |
| Remote worker | The assignment can run separately and return for review. | Credentials, network, environment and data retention. | Pinned environment + logs + reviewable branch. |
| Graph runtime | Persistent state and side-effect ordering dominate. | Replay, duplicate delivery and approval semantics. | State transitions + failure matrix + reconciliation. |
Local execution is not automatically private: model endpoints, telemetry, connectors and credentials still matter. Read the local-model guide.
03 / Session-only checklist
What goes with
the patch?
Check the artifacts you can actually produce. This is a handoff aid, not an audit, risk score or release approval. State stays in this tab and resets on reload.
0 of 8 evidence items marked. Completion is not approval.