Practice · Engineering capacity
How do I let AI agents ship faster without shipping more bugs?
Remove repeated investigation and verification work while keeping the acceptance standard explicit. Give the agent the relevant requirements, risks and prior failures before implementation, then carry the decisions forward and renew any evidence whose basis changes.
Make the second change start further ahead.
Illustrative workflow: the first importer fix establishes that malformed records must not partially update the store. It leaves a requirement, a failing-before/passing-after reproducer and a recorded decision. A later performance task retrieves that promise and reruns the relevant checks instead of asking a reviewer to rediscover it.
- Choose one component and a comparable baseline change.
- Recover the intent and settle only consequential ambiguity.
- Let the agent implement within that authorized scope.
- Execute the required checks at the candidate revision, inspect the results and keep gaps visible.
- Retain the results and measure what the next change reuses.
Proof is intended to reduce the human work around acceptable changes. It does not make the coding model itself faster, and we do not publish a measured time-saving claim for this workflow yet.
Measure time to an acceptable change, including setup.
| Measure | Baseline | Proof pilot |
|---|---|---|
| Setup and requirement recovery | Record active minutes | Record active minutes, including installation |
| Investigation and clarification | Record active minutes | Record active minutes |
| Review and verification | Record active minutes and reruns | Record active minutes and reruns |
| Rework | Record active minutes and repeated cycles | Record active minutes and repeated cycles |
| Elapsed decision time | Record start and disposition time | Record start and disposition time |
| Compute and tool cost | Record separately | Record separately |
| Later defects and reversals | Use the same follow-up window | Use the same follow-up window |
| Knowledge reused on a second change | Record artifacts actually used | Record artifacts actually used |
This is a blank pilot worksheet, not a customer result. Compare similar scope and consequence. Include rejected, blocked and deferred changes as well as accepted ones. Faster merging is not an improvement if the effort moves into setup or later repair.
Keep the bar while reducing repeated work.
Run relevant checks through your existing CI, propagate their failures and require current results before merging. Check that required jobs actually execute and cannot silently pass by being skipped. Approval follows the saved project policy; merge and deployment permissions remain separate. A passing check is useful only for the claim and revision it actually covers.
Read the complete workflow and agent guardrails. The US defense AI4SDLC guidance offers adoption and assurance practices; it is not evidence of a productivity gain from Proof.
Start with a change you need to trust.
Bring one component, its build path and a change you are not comfortable delegating yet. We can identify the intended behavior, missing evidence and decisions together.