The interesting question isn't whether engineers use AI. They do. The interesting question is how the workflow changes when half the team is shipping with Claude Code or Cursor and half isn't.
What works
- Spec-first. Write a one-page spec before asking the agent to implement. Vague prompts produce vague code; specs survive the back-and-forth.
- Plan, then execute. Have the agent produce a plan as a separate step. Review the plan. Then execute. Cheaper than reviewing a 500-line PR.
- Tight feedback loops. Tests, types, linters. The agent uses them too. A repo with strong CI gets disproportionately more value from AI.
- PR as the unit of review. No different from human PRs. Diff, description, tests. The author (human) takes responsibility regardless of who typed the code.
- Atomic commits. Smaller commits = easier review, easier reversion when the agent gets it wrong.
What to avoid
- Long autonomous runs without checkpoints. "Go build the whole feature, come back when done." The diff is unreviewable, and one mistake at step 3 poisons the rest.
- Skill creep. Adding every plugin and skill before knowing what the team actually uses. Start small, observe, expand.
- Hiding usage. Reviewers should know an agent wrote the code. Not for blame, for calibration. Different failure modes than human code.
- Treating it as a junior dev. It's not. It's a tool that's very good at some things, very bad at others. Calibrate your trust per-task.
Onboarding
- A
CLAUDE.md or AGENTS.md at the repo root explaining structure, commands, and conventions. The agent uses it. So do new humans.
- A shared set of skills, slash commands, or prompt snippets in version control. Reduces drift.
- Pairing sessions where a senior shows how they prompt, review, and reject AI output. The skill transfers fast once seen.
Where teams get stuck
- No evals on the actual workflow. "Does it work?" answered by feel. See Evals.
- Two cultures. Half the team relies on AI heavily, half refuses. Productive on day 1, friction on day 90. Pick a baseline.
- No cost visibility. Engineers burn $40/day in tokens and no one knows until the bill arrives. See Budgets.