Coding with an AI agent: what actually changed (and what didnt)

Short summary

A practitioner shares lessons from months of shipping merged open-source PRs with AI coding agents in the loop. The biggest win is codebase navigation, not code generation—asking an agent to explain unfamiliar code collapses hours into minutes. The critical discipline is verification: the author runs old and new code side-by-side and asserts identical output before merging. The core principle is delegation, not abdication—let the agent type, keep the thinking.

• •AI agents excel at codebase navigation and boilerplate, not autonomous problem-solving
• •Verification is the key 2026 engineering skill—reviewing AI output is now mandatory
• •Human stays at both ends: problem selection and final sign-off, agent compresses the middle

Generated with AI, which can make mistakes.