Tom Brackett

Replit13 August 2026

The only agent that could read its own logs

Brief six is a scheduled importer with a queue and a webhook receiver. Nine of the eleven tools in this index cannot express it. Replit finished it on two of three runs, and on the failed run the agent correctly identified the cause from the container logs before running out of context. That capability comes from being a real Linux box rather than a sandbox, and it is why Replit tops our scalability axis at 4.5. The counterweight is that the same brief produced our ugliest interface of the day.

Bolt.new13 August 2026

Wins the first hour, loses the third day

Bolt was fastest to a running app on five of six briefs, sometimes by a wide margin, because the toolchain runs in the browser with no build queue. It was also the only well-funded tool where two of my runs ended in an unrecoverable state: a broken dependency tree the agent could not unwind, where restarting was cheaper than continuing. Token consumption during those attempts was the least predictable I measured. A 3.7 on reliability against a 4.7 on speed, and the gap between those two numbers is the review.

Rocket.new13 August 2026

Real mobile output, shaky everything else

On the mobile brief Rocket.new produced Flutter and React Native targets whose navigation followed platform conventions rather than looking like a wrapped web page, which nothing else here managed. On the other five briefs it needed more regenerations per finished screen than Lovable or Base44, and our SEO axis measurement of 3.0 is the lowest in the index. One panellist argues our weights are unfair to a mobile tool. He is partly right, and the objection is published on this page rather than filed away.

It is. Lovable and Replit each had one unrecoverable run out of eighteen, v0 had none, and Emergent had four.