Agent-first engineering begins with human intent. Agents help translate that intent into systems through a deliberate sequence:
Human intent → specification → architecture → implementation → verification → evidence → human review → deployment.
Each stage produces something inspectable: requirements, design decisions, code, test results, deployment plans, and operational observations. Humans retain authority over consequential changes. Evidence makes that review concrete.
The sequence also needs feedback. A failed test may expose a design assumption. A deployment plan may reveal a missing permission or an unclear ownership boundary. An operational result may show that the specification defined success too narrowly.
Those findings should travel backward to the earliest decision that needs correction. I call this engineering backpropagation: implementation findings improve architecture, verification findings sharpen specifications, and deployment observations refine our understanding of the original intent.
The term describes an engineering feedback process, not automatic model training. Learning becomes durable when findings are captured in specifications, tests, architecture decisions, and operating procedures that future work can use.
Agents can help connect those records, propose corrections, and rerun checks. Humans review the evidence and approve deployment. Over time, the workflow can accumulate better constraints and fewer repeated mistakes.
An agent-first system becomes more useful when each completed task improves the evidence available for the next decision.