Approach
Engineering systems that can investigate, implement, verify, and correct work without surrendering control.
AI-native engineering is not prompt-driven improvisation. It is a software-delivery system built around repository context, bounded authority, deterministic quality gates, runtime evidence, escalation, and human accountability.
Deterministic systems where correctness is required, AI where interpretation adds value, and humans where judgment and accountability remain necessary.
What remains deterministic
Correctness, state, policy, and proof.
- Business invariants and state transitions
- Authorization and destructive-action controls
- Tests, type checks, build gates, and contract validation
- Source control, change isolation, and release provenance
- Evidence boundaries between local, preview, deployment, and production
What requires human judgment
Intent, tradeoffs, risk, and accountability.
- Ambiguous product requirements and strategic priorities
- Architecture decisions with long-term cost
- Security, privacy, confidentiality, and compliance decisions
- Irreversible operations and material scope expansion
- The final decision to merge, deploy, or release
Authority model
Delegate execution without delegating accountability.
What an agent may complete, what it must escalate, and which actions remain prohibited.
Agent authority and human responsibility across engineering activities | Activity | Agent authority | Human responsibility |
| Read code and investigate | Autonomous | Oversight |
| Clarify material ambiguity | Escalate | Decide |
| Implement approved scope | Autonomous | Review when needed |
| Run checks and previews | Autonomous | Oversight |
| Retry bounded failures | Autonomous within limit | Intervene after limit |
| Change authentication or schema | Implement and verify within approved scope | Approve scope and consequential decisions |
| Integrate candidate | Follow repository lane and granted authority | Own required decisions and gates |
| Release production | Prepare evidence; execute with separate authority | Own release decision |
| Bypass verification | Prohibited | Prohibited |
| Expose secrets | Prohibited | Prohibited |
Read further
The method, the practice, and the setup.
The method
The full operating model: lifecycle, accountable delegation, verification selected by risk, bounded correction, and release authority.
In practice
The same method applied to an offline-first point of sale, with physical-device verification and production boundaries.
The setup
Where the agents run, how the server is organized, and what running around the clock really means.