Coding agents are only as effective as the context they can see—and for long-lived software, the repository is their real prompt. This practical talk explores how clear specifications, enforceable constraints, observable acceptance criteria and fast feedback loops can turn repeated AI mistakes into improvements that compound over time.

A walkthrough of a real session where two developers and their AI agent teams collaborated on a complex architectural change — going
from a rough spec to 8 PRs, 4,500 lines of code, and a clean task board in a single day.

The talk covers how we used git worktrees as a shared context surface between two developers' Claude Code agents, spawned specialist
subagents (including a Redis expert that caught two critical production bugs during design review), ran parallel implementation
across multiple branches, and coordinated via a dedicated agent that polled for commits and monitored Slack and CI.

The focus is on the collaboration model — not the technical feature itself — and what the human's role becomes when AI agents can
read, write, test, and coordinate autonomously.

Our coding agents are getting smarter, but the workspace we drop them into is probably not enabling them to use their full potential. The decisions, failed attempts, conventions, and product knowledge are still crucial context scattered across people’s heads, Confluence, old conversations, and forgotten pull‑request comments.

During this talk, we'll walk through the shift from the “good old” prompt engineering to agent experience, then I'll show a practical AX experiment at my current company that led to a better workspace for both agents and the people using them.

They hired me to add features. The codebase had less than 1% test coverage. Azure was retiring .NET 8, but upgrading seemed too hard, so they'd waited years. The kind of problems you know you should fix but never do.

Four weeks later: 560 commits, 65% test coverage, 13 database improvements, and yes, we shipped features too. Two programmers, four hours a day, one $17/month Claude subscription.

Here's what nobody tells you about AI: Day one, it took us eight hours to fix one nullable reference. Week four, I did fifty in under an hour with five minutes of my time. We didn't get faster because AI got smarter. We got faster because we built the infrastructure—scripts, CLIs, knowledge docs—that let AI scale what we'd already figured out.

This is a field report on what's actually possible when you stop thinking "AI makes me faster" and start thinking "AI lets me solve problems that used to be too expensive to touch."

Come see how we got a codebase under control.