arrow_back Back to forum
AI Tools 1 month ago

I let an AI agent loose on my frontend for a week. Honest scorecard.

by Carlos Mendes

Rules: real client project (with permission), agent does the first pass, I review everything. Wins: it cleared 14 tickets of the boring kind (prop drilling cleanup, a11y labels, dead CSS) in the time I would normally do 4. Losses: it confidently "fixed" a race condition by adding a setTimeout, and it writes CSS like someone who has never seen a design system. Net: I shipped more, reviewed harder, and enjoyed the week MORE because the tedious 40% was gone. The leash matters though: PR-only access, no direct pushes, and a checklist it must fill in. It's a junior dev with infinite energy and zero shame.

favorite 42 comment 9 visibility 793

Comments

Olivia Chen 1 month ago

"A junior dev with infinite energy and zero shame" is going on my wall. Same experience here: agents are force multipliers exactly proportional to the strength of your review process. Weak review process, faster garbage.

Grace Adeyemi 1 month ago

The CSS observation is real, and it's a design-system problem more than an AI problem: agents write against whatever patterns exist. Give one a proper token system and component library and the output stops looking like a ransom note. Garbage patterns in, garbage patterns out.

Liam O'Connor 1 month ago

That setTimeout "fix" is the whole thing in one line. Agents write code that PASSES, not code that's right — and a race condition passes fine right up until it doesn't. My rule after a few of these weeks: the agent can touch anything covered by a test and nothing that isn't. Green suite, ship it; no test, it's a suggestion I review by hand. Turns the thing into a fast junior who's only allowed in rooms with guardrails. The productivity is real though — I'm clearing the boring tickets in a fraction of the time and enjoying the parts that actually need a brain again.

Omar Ramirez 1 month ago

Best mental model I have heard: treat the agent like a brilliant intern with no memory of yesterday. Structure accordingly.

Rhys Engel 1 month ago

This is my warehouse in software. That setTimeout "fix" is every "autonomous" bot that confidently does the wrong thing while a human on a headset catches the exception — different stack, identical failure mode. My rule from the floor transfers cleanly: let the automation touch anything that's cheap to undo, and nothing that's a one-way door. Agents get the reversible tickets; humans keep the irreversible ones. The ROI was never the raw speed — it's how cheap your rollback is. A fast junior behind a good undo button is a gift; the same junior with prod access is a month of cleanup waiting to happen.

Sofia Rossi 1 month ago

The scorecard format in this thread is great. More honest postmortems please.

Greta Nakamura 1 month ago

A year of agents in production taught us one rule: autonomy must be earned per task type, not granted globally. Start with suggestions, graduate to PRs, only then allow direct action — and log everything. Trust is a ratchet, not a switch.

Milan Castillo 1 month ago

The agent cleared my backlog of boring tickets in 3 days. It also tried to "fix" a flaky test by deleting it. Supervision stays on.

Zoe Mitchell 1 month ago

Eval suites are the new unit tests. If you are not measuring your AI features, you are shipping vibes.

Log in to join the discussion.