Software factories: build or buy
A protocol for quality, human time and total cost
Slide contents
1. Software factories: build or buy
A protocol for quality, human time and total cost
2. A winner must wait for acceptance
Contract first, then attempts, then a decision
3. Activity is not yet impact
Adoption → flow and time → economics
4. The tasks probe different work
A bug, a feature, a migration and evolving requirements
5. T4 inherits its own T2
Shared base for independent tasks; own history for evolution
6. The evaluator owns the verdict
Requirements are public; checks are protected from worker writes
7. Compare product-and-model bundles
A plain agent session reveals orchestration overhead
8. Stopping rules come first
60 minutes and at most two external repairs per task
9. Failures stay in the denominator
Accepted, rejected, stopped, blocked or unrun
10. A faster agent may not shorten work
Execution, waiting and human effort are separate
11. Every attempt costs money
Accepted-task cost includes failed runs
12. A cheap run can become costly
Our arithmetic uses invented amounts, unrelated to any product
13. Human minutes change economics
Report time first, then state an assumed rate
14. A small experiment cannot rank a market
20 nominal attempts; repeating the full set gives 40
15. A green result has an observation window
Separate functionality, later changes and operations
16. Compare the work you still own
A subscription, a self-hosted system or a plain agent
17. Conclusions need an evidence package
Fixture, logs, acceptance, bills and failure recordings
18. A factory must justify its overhead
Accepted outcomes, work flow and complete cost