What if an AI could write your code, but only if your organization rewires itself?
The Simple Version
Think of AI agents as a new kind of employee that needs a desk, tools, and a manager, not just a prompt.
How It Actually Works
Agents Need Harnesses
An AI agent works only when you wrap it in a technical harness that handles input, output, and error paths.
From Prompting to System Design
Instead of chasing perfect prompts, you design loops that let the agent call tools, read logs, and trigger fixes.
Technical Deep Dive
Patrick tracks two numbers: human touches per result, which should drop, and the shareability of each fix, which should rise.
He moves planning from vague backlogs to tight slices that can be handed off, while keeping larger conversations separate.
Above the team, a platform team provides paved roads and a named owner, preventing a thousand ad‑hoc experiments.
Why This Matters
When harnesses become reusable, one improvement benefits the whole org, turning a single 10x boost into a collective gain.
Teams that stop fixing agent output and start improving the surrounding system see fewer regressions and clearer ownership.
This shift mirrors the early days of continuous delivery, where the technology was ready but the culture had to catch up.
Key Takeaways
- Agents need technical harnesses, not just better prompts.
- Measure human touches and reuse to drive improvement.
- Shared harnesses multiply impact across the team.
- Paved roads and a named owner replace chaotic experimentation.
- Planning splits into small hand‑off work and larger conversation work.
- Retros should focus on where the agent hits walls, not on code bugs.