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

Every job post now says "we expect you to use AI in your daily workflow." Good — because used well, an AI coding agent is a real multiplier. Used badly, it's a confident intern that ships bugs. After months of building and shipping merged open-source PRs with an agent in the loop, here's what actually changed — and what didn't.

What genuinely got faster

Navigating unfamiliar codebases. Dropping into a repo I've never seen and asking "where is X handled and why" collapses hours into minutes. This is the single biggest win — not writing code, but understanding it.

Boilerplate and tests. Regression tests, argument parsing, config plumbing — the stuff you know how to write but don't want to. The agent drafts, I review.

Refactors with a spec. "Replace this nested loop with a vectorized construction, preserve ordering and the guards" → a clean patch I then verify.

What did NOT change (and never should)

You still own correctness. I don't merge anything I can't explain and haven't tested. On one fix I had the agent generate a change, then I ran the old and new code side-by-side across every input and asserted identical output before opening the PR. The agent wrote the diff; I signed off on it.

Judgment is still yours. Which issue to pick, whether the scope is sane, whether a bounty is a dead race — the agent will happily execute a bad plan. Deciding what to do is the human's job.

The workflow that works for me

1. Human: pick a small, concrete problem.
2. Agent: explore the codebase, propose a minimal change.
3. Human: sanity-check the plan (reject vague or oversized ones).
4. Agent: implement + draft a test.
5. Human: run the full suite, verify, read every line.
6. Human: write the PR, disclose AI assistance, ship.

Notice the human is at both ends. The agent compresses the middle.

The honest failure mode

The trap is trusting output you didn't verify. An agent will produce plausible, well-formatted, wrong code — and it looks exactly like the right code. The discipline that matters in 2026 isn't prompting; it's verification. Reviewing AI output is now a core engineering skill, not an optional one.

Delegation, not abdication. Let the agent do the typing. Keep the thinking.