---
title: 'In the News: November 17, 2026 (Extra 3)'
description: "A StarSling CI agent fixed a quadratic cleanup bug in better-auth's tests and shipped the fix with a 400-run A/B benchmark. It was merged."
canonical_url: 'https://darkfactory.dev/news/2026-11-17-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-17-extra-3.md'
collection: news
date_published: '2026-11-17T23:02:00-05:00'
date_modified: '2026-11-17T23:02:00-05:00'
---

# In the News: November 17, 2026 (Extra 3)


An agent from StarSling, a new CI company, found a bug in better-auth's shared test harness that had been slowing its adapter test jobs since September 2025. The pull request it opened came with a controlled benchmark, and a better-auth maintainer merged it.

## 1. A CI agent's pull request arrives with its own controlled experiment

**[perf(test-utils): stop re-deleting cleaned rows](https://github.com/better-auth/better-auth/pull/10762)** · starslingdev[bot], StarSling · better-auth pull request #10762 (merged)

The pull request explains that better-auth's adapter test harness deleted the rows each test created but never forgot them, so each cleanup re-issued deletes for every row earlier tests had made. On a 455-test battery, it says, 98% of delete calls (101,366 of 103,554) targeted rows that were already gone. A catch block swallowed the "not found" responses, "so the only symptom is a slow job." The fix moves one delete statement out of an inner loop. The pattern dates to a September 30, 2025 rewrite (#4844).

The bot ran both versions interleaved in a single run, 50 replicates each, for 400 job runs with zero failures. Measured on vitest's reported duration, mongo-adapter fell from 24.3s to 4.5s (5.4×), kysely-prisma-adapter from 145.6s to 45.7s, prisma-adapter from 287.0s to 170.6s, and drizzle-adapter from 185.5s to 156.7s. All four 95% confidence intervals exclude zero, and every run reported the same test counts as its counterpart, so none of the saving comes from running fewer tests. A contributor, dreamorosi, measured independently on a DynamoDB adapter: the generated conformance file fell from 88.23s to 11.81s. Maintainer bytaesu merged the change after review and said he would "periodically list and check Starsling bot PRs, like Dependabot!"

StarSling's founders introduced the company in a [Launch HN post](https://news.ycombinator.com/item?id=50536468) (62 points and 35 comments at 23:01 EST, November 17). Their harness has agents analyze workflows, run logs and machine telemetry, propose a change, run it beside the unchanged workflow, "deterministically judge that the difference is statistically significant", and ship the winner as a pull request with the benchmark attached. The company's other figures are its own and were not checked here: Mastra's slowest workflow cut from 29m 56s to 5m 06s, and an average 10x wall-clock speedup across five open-source benchmark repos.

**Why it matters:** An agent can open pull requests far faster than people can review them. This one gave the reviewer what the decision needed: the mechanism, a small diff, a replicated A/B run with intervals, and a parity check on test counts. If your agents open pull requests, that is a useful bar to hold them to.
