The Engineering of Intent, Chapter 11: The Art of Agentic Debugging

This is Part 11 of a series walking through my book The Engineering of Intent. In the previous chapter, we walked through the five-layer quality gate stack that catches the structural problems. This chapter is about the bugs that slip through — the ones that don’t reproduce locally and whose causes are spread across hundreds of files you didn’t personally write.

Debugging in the AI-Native Regime Is Archaeology

In the AI-native regime, debugging is archaeology. The code may have been written by an agent you supervised loosely. The commit message may not describe what you were actually thinking that Tuesday afternoon. Chapter 11 gives you the tools to reconstruct causation after the fact — which is most of the job now.

The Self-Correction Loop

Reproduce, capture full trace, hand to the agent, ask for hypotheses and diagnostics, iterate. The loop converges in a few iterations when the human insists on root cause before accepting fixes. The loop fails when the human says “just fix it.”

Bisection, Observability, and the Caching Heisenbug

Two disciplines the chapter argues are non-negotiable:

• Learn git bisect cold. Scripts that automate bisection on common failure modes pay for themselves within a month. In AI-native codebases, the offending commit is often second-order; bisect to the commit, then use the agent to trace the true cause.
• Observability as substrate. Agentic debugging is only as good as the traces you can feed. Structured logs, distributed traces, state event histories. If you cannot produce the full sequence of events for a reported bug in one query, fix observability before you fix the bug.

“An intermittent bug for one customer. Hours of manual investigation turned up nothing. The agent, given the request history and cache configuration, hypothesized an edge case in cache key computation conflating users whose IDs differed by trailing whitespace. Reproduced in fifteen minutes, fixed in twenty more, and codified a ‘trim at the boundary’ convention that made the class impossible.”

The Three Debugging Anti-Patterns

• The just-fix-it anti-pattern — accepting a fix without confirming the root cause.
• The context-overflow anti-pattern — adding more files when the question itself is malformed.
• The swap-the-model anti-pattern — reaching for a different model instead of restating the problem.

Next up — Chapter 12: The GenDD Pod. Part V of the book zooms out from the individual and the codebase to the team. Chapter 12 introduces the GenDD Pod — the team structure that actually ships AI-native work sustainably — and the handful of roles whose job descriptions have changed the most in the last two years.