A public technology company asked us to train its engineering organization on agentic development. Not a tiger team. The whole organization. Their diagnosis was sharp: every engineer already had an AI coding tool, and no two of them used it the same way.
The problem was not adoption
Adoption was near total. The problem was variance. Some engineers treated the tool as autocomplete. Some handed it entire features and merged what came back. Code review had no shared standard for work an agent had written, so reviewers either rubber-stamped it or rewrote it.
Speed went up in places. Confidence went down everywhere.
What the curriculum covered
Prompting as specification. The quality of agent-written code tracks the quality of the brief. We taught engineers to write the acceptance criteria before the prompt, the same way they would for a colleague.
Review as the job. When an agent writes the code, review is where the engineering happens. We rebuilt the review checklist for agentic work: provenance, test coverage the agent claimed versus the coverage that exists, and the failure modes agents reliably introduce.
Guardrails in the workflow, not the person. Branch protections, required tests, and gates that stop an agent's change from reaching main without a human decision. Standards that hold because the pipeline holds them.
Accountability that does not move. The engineer who merges owns the code. Whatever wrote it.
How we measured it
Not seats. Not hours saved on a survey. Shipped work: pull requests merged per engineer, defect rate on agent-assisted changes, and review time. Adoption measured in output is the only kind that survives a budget review.
What changed
Standards where there had been habits. Review that caught what the tools got wrong. And an engineering organization that could hire for the new way of working because it could finally describe it.