Making AI part of a controlled engineering system instead of an unreviewed shortcut
Teams wanted faster delivery from large language models, but token usage and code generation alone could not provide architectural consistency, acceptance-criteria coverage, or accountable decisions.
The decision
Give each agent a narrow responsibility and place human review and approval between meaningful transitions.
Observed result
Developers could iterate plans before coding, compare implementation against tickets, create stronger PR context, and diagnose cross-stack bugs in the same day in situations that had previously consumed several days. Results varied by task and team.
The decision, visualized
AI speed + human gates
Narrow agents supported the work while developers and product owners retained decisions.
The system
The workflow separated plan creation, senior-level plan review, implementation, self-review against the ticket, pull-request creation, documentation, and independent PR review. A developer could repeat planning until the approach made sense rather than accepting the first plausible answer.
Human control
The agents could flag missed criteria or propose extra behavior, but a developer and product owner decided whether that behavior belonged. The system was designed to keep the developer in control rather than let an agent run indefinitely through the codebase.
Cross-functional documentation
PR descriptions included high-level changes, the developer’s happy-path checks, and likely regression areas for QA. Related documentation work made test users, application behavior, and current implementation context easier for remote teams to retrieve.
The operating belief
More tokens do not automatically produce a stronger team. AI must improve learning, review quality, documentation, and feedback while preserving the judgment the company hired its people to provide. The economic and data consequences belong in the design too.