Putting AI into practice: coaching, context and verification

Buying coding agents is easy. Getting engineers to use them well took hands-on leadership, shared context and verification that never depends on trust.

Case study. I set the direction and coached; others on the team built parts of the documentation automation and training.

The situation

By 2025, coding agents were capable enough to change how software gets built, and most of the work in front of my teams was in existing systems whose rules the code does not fully explain. Licenses alone were not going to change outcomes.

What I did

Worked hands-on with agentic tools from 2025, including Claude Code, Codex, GitHub Copilot and Cursor, so what I asked of teams came from practice.
Coached individual contributors in spec-driven development, AI-assisted review and automated validation.
Built the context agents need in existing systems: tools, connectors and codebase knowledge, so an agent starts from how the system actually works.
Asked teams to make their learning reusable, feeding what worked into shared instructions, prompts, workflows and guardrails.
Kept verification ahead of speed: automated tests, including AI-assisted Playwright tests, and proof of validation before QA.

What changed

Shared context and instructions now live alongside the code, so agents and new engineers start from the same knowledge. In specific cases, new engineers reached their first production change much sooner than in earlier examples.

What’s next

The next step is APEL, an agentic product engineering lifecycle designed for agent-assisted work across all aspects of the delivery lifecycle.

What I would carry forward

Tools spread fastest when leaders use them too, when context is shared rather than personal, and when verification never depends on trusting the output.