AI + Playwright is turning browsers into agents. 5 patterns from 20 open-source underdogs 👇

Every hour saved by AI coding tools is lost again if your QA isn’t automated and self-healing.

Just integrated Vercel AI Gateway at Bug0. I think it’s pretty great for a couple of reasons: - no extra markup - you can BYOK - in case of failures, they automatically retry using their system creds so you get extra reliability and much higher rate limits - send

Always bet on LLMs to get better. When we started Bug0 ~3 months ago, I wasn’t sure if letting AI directly test web apps was the right decision. To get determinism, we started by generating stable locators and building AI-powered auto-healing. Fast forward to today: we’ve

We just crossed 10K test executions at Bug0 in the last 3 months. My biggest learning in getting LLMs to do what you want: - LLMs are non-deterministic tools. Do not over-optimize trying to make them deterministic. - use tool calling to do tasks deterministically. - use

playing around with this CTA label… thoughts? 👀

Once you do this, you cannot go back!

🆕 Every week, Michelle Fang 🌁 tracks dozens of developer meetups in SF. There was a Thursday last year with 10 (!) AI meetups. It's really hard to stand out. dx.tips/hosting Sarah Chieng guests with how she gets over 1000+ RSVPs and way more online) for her events:

If one of your angel investors berates you into using their product, they aren't going to be there for you when times are hard. The first thing I tell founders I invest in is: always do what's best for your business, get advice, but go with your own gut.

when i was building the v0 of Bug0 AI agent, in the early days our AI agent sometimes wandered off and started testing competitor websites 😂 AI does weird stuff have guardrails from day 1

I have found that Grok Code Fast 1 (grok-code-fast-1) is a very capable model for multi step agentic loops. You can achieve complex tasks without breaking the bank. but `gemini-2.5-flash` is way faster, in some cases up to 50% faster (it costs $0.10 extra per million tokens).

We get ~1k pageviews a day on our new Bug0 site. Traffic isn’t the issue. Indexed on Google. Showing up in AI summaries. The issue: conversions. CTAs at 6% (good for B2B) but only 1–2 demos booked. Some days, none. Why? Nobody reads anymore. Landing pages all look the

since we started building Bug0 we’ve seen so many moments where a bug was just about to be shipped… but Bug0 caught it. why does this work? traditional browser testing is broken for one simple reason: fast moving teams ship so fast that the tests are outdated the very next

In long agentic loops, adding a simple self-check like `doesActionAdvanceUsTowardsGoal` massively improves results. The AI takes fewer wrong turns, and when it does, it often flags false, giving you a chance to catch it and retry the step correctly.

recently Bug0 AI caught a bug that regular playwright missed. instead of calling .clear() and .fill(), it acted like a real human - pressed cmd + A then typed. ..that subtle difference exposed the bug. browser tests need more than automation. They need to mimic humans!

Today, we’re launching Cloudflare Email Service. Send and receive email directly from your Workers with native bindings—no API keys needed. cfl.re/4mx6fbO #BirthdayWeek

We're wasting our most talented engineers on a fundamentally broken process. We ask them to build innovative products, then saddle them with writing and maintaining brittle E2E tests. Every hour an engineer spends fixing a cy.get() selector or a flaky Playwright test is an hour

Exciting!

Super impressed with Codex so far. I’m using the Codex plugin in VS Code, and gpt-5-codex-low has been more than enough. This is the first time something’s actually outperformed Claude Sonnet 3.5/4 for me. If it keeps working this well, I might just cancel my Copilot and