From ad-hoc prompting to agentic engineering

On a small engineering team, AI work had drifted into prompt-heavy, one-off sessions. We coached people on how they already worked, then rebuilt the practice around that: agentic workflows with clearer steps, repository guidance so models and humans shared the same context, and evaluation so the team could tell whether a change helped. Measured by features shipped per sprint, output roughly doubled, and longer-serving engineers grew into ownership of demos, standards work and critical paths. The same pattern of coaching, structure and validation later transferred beyond engineering, where the harder problem is knowing when an ambiguous answer is actually correct.

What we delivered

• Coaching that turned ad-hoc AI use into structured, agentic processes with repeatable quality
• Hands-on evaluation of coding agents on real product work, with findings shared back to the team
• Repository readiness for AI-assisted development, including agent guidance and LLM-friendly documentation
• A validation habit: expected inputs and outputs, marking criteria, and regression checks when practices changed