From harness to software factory
What turns individual runs into a durable flow of work
Slide contents
1. From harness to software factory
What turns individual runs into a durable flow of work
2. A factory manages a flow of work
From one run to state, checks and acceptance
3. The harness sits inside the factory
The model chooses an action; the harness executes; the factory manages work
4. Each task needs a completion contract
Intent, starting revision, checks and permitted output
5. Autonomy has a concrete endpoint
A diff, CI, PR, merge and production behavior are different results
6. Sessions end; tasks remain
Persist intent and evidence outside the worker’s memory
7. Required steps surround the agent loop
A deterministic frame can contain flexible implementation
8. Minions reuses the developer environment
An isolated devbox, context, tools and a PR for review
9. Retries need an explicit limit
Stripe: local feedback before full CI
10. Honk changes code inside Fleetshift
Migration orchestration existed before the agent
11. Consistency reduces uncertainty
A catalog, documentation, standard environments and fast feedback
12. Both systems preserve work management
Compare responsibilities, not generated line counts
13. A tracker can feed the factory
Symphony turns queued work into isolated runs
14. Every transition has an owner
People define the contract; the system executes; acceptance decides
15. The queue also shows stopped work
Ready, running, needs a decision, accepted
16. The fixture makes effects observable
An event crosses the API, SQLite, delivery and a React history view
17. Recovery starts by reconciling state
Stop after the code changes, before the final report
18. A result needs an inspectable trail
Task, commit, checks, decision and cost belong to the same run
19. Practice needs two recorded paths
Normal completion, then interruption and recovery
20. Start with one manageable task
A queue, durable state, checks and an explicit boundary