Workflow

Agentic development starts with building its environment with a well-defined ruleset and clear boundaries, because that is what makes the output more predictable.

I fix the rules and boundaries the agents must work within: best practices, team and syntax priorities, expected style, toolset, frameworks and libraries, an agent-by-agent review process, and task-specific agents and skills.

Before any code is written, I work the problem through with the AI in plan mode.

Together we produce a detailed plan covering what the goal actually is, what could go wrong, what I failed to consider while planning alone, what I judged incorrectly, and which pitfalls are likely to surface during implementation.

Implementation is agentic, bounded by the ruleset of the environment, which governs how we want to achieve our goal.

Automated tests written alongside the code, just as I would do if I were developing manually.

The hardest part, and my main focus, is that agents do not necessarily know (and sometimes must not know) the domain-specific and business requirements. That easily breeds misunderstandings and implementations that are fine in general, but do not achieve what we actually want here.

After implementation, the review and any resulting fixes are also carried out by agents, before the code ever reaches human eyes for the final word.

Agents catch each other's blind spots: what one produces, another reads without attachment to it, so most defects surface while a fix still costs minutes rather than a full review round. By the time I look, my attention goes to the decisions that genuinely need a human, not to issues a machine could have caught.

Under no circumstances do we ship agent-generated code without a human reading it critically.

The responsibility still rests with the person who released it, however that code was produced. And if you do not understand what you are handing over, that is worse still.