Javi Murillo

The secret ingredient to Flaky Tests? Bad data. Stop blaming bad tests. Start focusing on how you seed your data. The success of your QA process depends on it.

14

1 Comment

Kernel aims to support every protocol that an AI agent needs to do its work. Today we’re introducing support for WebDriver BiDi, a W3C open protocol that allows agents to access and navigate websites. BiDi enables first-class support for Vibium, a new browser agent framework from the creator of Selenium. We’ve partnered with the Vibium team to build a native integration with Kernel browsers, which you can try on our platform today. We'll continue enhancing our platform to provide bleeding edge support for AI agents to access the internet. Give Vibium x Kernel a try and let us know how it goes!

28

1 Comment

Great post by Aleix Morgadas and Eugene Sergueev on how Flo Health Inc. revamped mobile onboarding If you’re keen to learn how to: 🔹 break out of release-cycle bottlenecks with server-driven architecture 🔹 scale experimentation without scaling engineering effort 🔹 build internal platforms that empower teams across product, design and content …it’s well worth a read: https://lnkd.in/emiqM7NM

6

Spotify’s use of AI to automate large-scale engineering work highlights how agentic tools are already reshaping how enterprise software is built, maintained, and deployed. When you pair that with OpenAI saying GPT-5.3-Codex was instrumental in creating itself, it really drives home how fast this space is moving.

4

If you’re a CEO/founder or PE operator and your tech org is: ❌ Shipping late ❌ Stuck in reactive mode ❌ Struggling through the Series A/B jump, or post-acquisition integration or after a major release It’s not “a delivery problem.” It’s an operating model problem. At the inflection point, the old playbook breaks: what worked at 10 engineers fails at 40… and breaks again across multiple products/teams. That’s where I come in as a fractional CTO: - Spot platform and delivery risks before they hit revenue and retention - Rebuild execution rituals (prioritization, ownership, release confidence) without derailing the roadmap - Align architecture + team structure to your value-creation plan (growth, margin, or exit timeline) If you’re feeling the strain of growth, I can help you turn “busy and reactive” into predictable shipping, and a platform that scales. #CTO #SaaS #StartupScale #PrivateEquity #TechStrategy

8

3 Comments

PostHog is perfect for engineers. PMs need a different lane. PostHog nails what engineers care about. But sticky adoption comes from deep hypothesis testing and measurement - work that lives squarely with PMs. Engineers should be shipping. Context-switching between code and funnels is a tax. 👉 Our take: -Engineers wire up the data. - PMs own the hypotheses - what’s working, what isn’t, and why. - That job can easily be full-time. 🌟 So we’re building a PM-native layer on top of PostHog: - Write a hypothesis → instantly pull the sessions/users that prove or disprove it - Measure intent, hesitation, and friction (not just clicks) - Trigger chatbots/pop-ups at the right moment - Map replays to customer conversations for fast follow-ups PostHog stays the engine. We’re giving PMs the perfectly handcrafter flow for hypothesis testing around adoption If you’re a PM using PostHog today, what’s the one workflow you wish existed?

4

Curious: has anyone found a particular sub-agent role or setup that's been especially powerful? Would love to compare notes. I've been experimenting with Claude agent teams lately, and one thing stands out: the software paradigms we've always relied on (API design, test coverage, modularity) matter more when working with AI, not less. Giving each sub-agent a specific role (Team Lead, Frontend Dev, Backend Dev, QA Engineer) produced noticeably better results than asking one agent to do everything. Wrote up what I learned using a calculator app as the example. Link in the comments.

12

1 Comment

The biggest QA problem I keep seeing isn’t execution. It’s the lack of a single source of truth. PRDs, Figma, tickets, people’s heads… scattered everywhere. Real confidence doesn’t come from running more tests. It comes from knowing what truly needs to be tested.

34

1 Comment

Our team at Electra Vehicles, Inc. is far too humble. And lucky for you, I got permission today to share some of the wildest and most exciting work that we do. Among the State Of The Art (SOTA). For example, in January we achieved less than 0.5% state estimation error (max absolute error) for a challenging chemistry: Lithium Iron Phosphate, in freezing, and scalding hot temperatures. We used hundreds of batteries in the lab and many broke under extreme conditions. (Most survived!) LFP dominates electric vehicles and energy storage industry. And, it matters. Every 1% SoC improvement translates to millions of dollars in Gigawatt scale energy storage and millions of tons of CO2. MILLIONS! I'm proud to say, this team outcome was one of the crowning achievements of my scientific career. (I'm just one of many science contributors at Electra Vehicles, Inc., and special shoutout to Revathi Soundarrajan, on this work) No, there was no AI involved in the design. And particularly it's because SOTA problems are not suitable for AI use. Take a look at this explanation from Harvard about how AI usage is effective within a boundary called the "Frontier", a growing and jagged wall expanding through the multi-dimensions. And SOTA technologies are by definitions always found beyond this boundary. No easy picking. Theres a far simpler and well known method that can guide you to challenge the status quo, dozens of books have been written on the topic. MrBeast uses it and it's mentioned in the book 10x is easier than 2x by Dan Sullivan. If you're interested, I'd start with this book! News: https://lnkd.in/ew4A7VDY Harvard: "Cyborgs, Centaurs and SelfAutomators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise"

23

3 Comments

Week one with Claude Code was rough. By week two, something flipped: CLI tooling became a native part of the workflow, and plan mode gave me structure without slowing me down. This is 𝗽𝗼𝘀𝘁 𝟮 𝗼𝗳 𝟯 I’m sharing this week about my Claude Code experience. ➡️ 𝗖𝗟𝗜 𝘁𝗼𝗼𝗹𝗶𝗻𝗴 𝗯𝗲𝗰𝗮𝗺𝗲 𝗮 𝗻𝗮𝘁𝗶𝘃𝗲 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄 Our projects lean heavily on custom CLI tooling we have built over time: test runners, linters, build scripts, deploy commands. In Cursor, getting the agent to use those tools was painful. I had to run commands manually, copy and paste the output, trim it down, and then re-feed the context back. In Claude Code, I can run commands in the thread just like I would in a terminal, and the agent remembers that for later. No more wasting context cycles explaining which command to run or pasting outputs back in. The shell and the agent finally feel like a single surface, which makes our whole CLI stack actually usable in practice. ➡️ 𝗣𝗹𝗮𝗻 𝗺𝗼𝗱𝗲 𝗴𝗮𝘃𝗲 𝗰𝗼𝗻𝘁𝗿𝗼𝗹 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗼𝘃𝗲𝗿𝗵𝗲𝗮𝗱 In Cursor, we relied on a planner prompt jammed into workflow.mdc. That meant everything got forced into a plan, even trivial edits, and I had to drag that setup across projects. Claude Code’s plan mode collapsed all that. I can call it when the task is complex, and it lays out steps, files, and validations. When it is small, I skip it. No hacks, no configs. The need to juggle models faded too. I do not think “Opus for reasoning vs smaller for edits” anymore. Plan mode when it matters, straight edits when it doesn’t. Best part: after a plan is approved, Claude Code auto-applies changes that map directly to that plan. Then it switches back to asking permission before new edits. And when running natively inside Cursor, those changes show up directly in the IDE’s diff view. It feels intentional, with planned steps mapped cleanly to code. My workflow now •  Use plan mode for multi-step tasks, skip it for quick edits. •  Run our CLI commands directly in-thread so logs and reasoning stay together. •  Trust the plan → run → refine loop instead of micromanaging models. This was the inflection point. Claude Code stopped feeling like a handicap and started feeling like a workflow. We were not patching over gaps anymore, we were actually faster. Tomorrow (𝗽𝗼𝘀𝘁 𝟯/𝟯): the experiments I tried on top of this flow, and what stuck.

12

7 Comments