Agentic Coding, Honestly Reviewed

The setup

One TypeScript monorepo, 180k lines, forty tickets ranging from copy tweaks to a database migration. Each agent got the same tickets, the same test suite, and the same rule: a change only counts if it passes CI and survives review.

Where agents shine

Mechanical breadth. Renaming across hundreds of files, writing missing tests, migrating deprecated APIs, translating a component library to a new token system — agents finished in minutes what would have cost an engineer a day of tedium.

They are also excellent readers. Asking an agent to explain an unfamiliar subsystem before touching it is now the fastest onboarding tool we have.

Where they still fail

Ambiguous product intent. When a ticket implied a business rule that lived only in a colleague's head, every agent confidently invented one. Failure was never loud; it was a plausible, well-tested implementation of the wrong thing.

The fix is unglamorous: write the intent down. Teams with strong specs got dramatically better agent output than teams with strong prompts.