Anish Mehta

What if coding agents could reliably implement your software? Not just for a single session... but for an entire project 🤩 I kept hitting the same wall: the agent would start strong, then drift, then hallucinate requirements I never mentioned. By session three, we'd lost the orientation completely. The problem wasn't the AI. It was me. I was "vibe coding"—prompting on instinct without structured constraints. So I explored a different approach: Spec-Driven Development. The idea: make specifications—not code—the primary artifact. You define what and why before the agent writes a single line of code. The specs become a "memory bank" of your intent. The agent references the memory bank and implements your intent. I tested three open-source frameworks (Spec-Kit, OpenSpec, AgentOS) to explore how Spec-Driven Development works. My takeaway? The specific tool matters less than the approach. Full breakdown of my learnings and experiences on my blog: https://lnkd.in/ek93HPnn #AICoding #AgenticAI #SoftwareEngineering #ClaudeCode #SpecDrivenDevelopment

27

22 Comments

I love the idea of giving more autonomy and agency to our non-dev contributors. Product managers are already able to create prototypes with AI. Now, they can go one step further and actually contribute to the codebase. Chatting with GitHub Copilot directly from Teams, or opening a pull request from an issue with Copilot — this is just the beginning 🚀 This will clearly impact our development workflows: 🤏Shorter lead times for small and medium tasks 💨Faster path from idea to production 🚀Fewer handoffs and less friction Of course, developers may feel concerned — maintaining additional code written with AI, even after review, can be uncomfortable. But this is also a great opportunity: 👉 to better document developer expectations 👉 to formalize coding standards, architecture rules, and best practices 👉 so the AI can follow those guidelines and produce compliant, maintainable code Empowering PMs doesn’t mean replacing developers. It means raising the bar for collaboration between humans and AI — and between product and engineering. Curious to see how this evolves 👀

3

I have asked this question to 20+ engineering leaders. ...not one had a ready answer. "What's the cost of 1 PR in your team?" I repeat "What's the cost of 1 {merged} PR within your engineering org?" The actual number will surprise you. A typical dev raises ~10 PRs in a sprint ...not all get merged. On a monthly basis that's 20 PRs at best. Cost of a developer {offshore} is say $35/hour You’re already at ~$300 per PR in dev time alone. Now you add Product Owner, Scrum master, testers. You are now staring at upwards of $500 per PR Now factor in reality ...waiting on code reviews, rework due to bugs & CRs...the effective PR merge rate is much lower. Welcome to the $1000/PR club. 🔥 With AI entering the scene, developers are raising more PRs But the bottleneck hasn’t disappeared. It's shifted to slow code reviews, poor test automation and & half baked requirements That's becos' we are shipping with yesterday’s playbook Engineering doesn't need more developers It needs a new O-p-e-r-a-t-i-n-g Model We’re building this new operating model in public. If this resonates, let’s talk.

36

3 Comments

If you’re going to give engineers homework in interviews, pay them for it. That’s been my long-held belief. But it’s more than just compensation: We’ve been thinking about experimenting with a “hands-on day” interview process at Depot. It’s coming from seeing how we are growing and the things we want to be better at, while also protecting the things that make us really good. It’s not an original idea. Several companies are starting to do one day work projects with candidates to get a feel for how they build, think, and communicate. They pay candidates a flat rate, something like $1,000 to work with the team for a day. It’s basically a real-world trial. You see how someone collaborates, debugs, and asks questions. You can interact more in real-time rather than someone working on a homework assignment in isolation. Candidates also get to see what it’s like to actually work in the company as well. It’s a two way street. Folks are generally far better in their jobs if they actually enjoy what they are working on and who they are working with. When you invest in this you’re also signing up for a higher standard of responsibility. You owe folks real feedback when they are doing homework assignments or working alongside you for a day. If they go through all of that and you just send a generic “no thank you email”, you’re an asshole. It’s about real feedback. We owe them a thorough code review and a real sense of what it’s like to collaborate with us, just like they were already part of the team. This approach requires significant planning and process to make it successful. Otherwise, it’s just a waste of everyone’s time. We’re not there yet with deploying this idea for Depot, but it’s definitely something we are looking at as we start hiring more folks. It will be a lot of work to get off the ground. But the insights we’ll gain into collaboration, problem-solving, and culture fit are invaluable.

31

2 Comments

Is it just me or does the vibe coding movement feel a lot like the no-code movement from 5-10 years ago? Same pitch. Anyone can build software. No engineers needed. It was supposed to democratize software. Forbes called no-code "The Most Disruptive Trend of 2021." Gartner predicted citizen developers would outnumber professional coders 4 to 1. People built all sorts of things with it. Tracking bee colonies at a honey company. Managing the snack rotation at a 50-person startup. An inventory system for a guy who sells vintage typewriters on Etsy. You know, the really important stuff. No-code ended up being helpful with the long tail of business software that nobody was ever going to build. Stuff too niche for a SaaS product to exist, but too annoying to get IT to prioritize. Now vibe coding is running the exact same hype cycle. So much feels familiar — the pitch, the analyst predictions, the "engineers are done" narrative. The differences are real though: vibe coding outputs actual code, there's no platform ceiling, and the code is actually portable. But what am I actually seeing people build? Mostly prototypes. Projects IT doesn't have time for. The same long tail no-code served — just with more headroom. Maybe this time really is different. Maybe vibe coding breaks out beyond the backlog and starts eating into serious engineering work and we will see vibe coded app being deployed to production and scaled to millions of users. Or maybe you'll still need an experienced engineer for anything more complex than a bee colony tracker. Time will tell.

71

27 Comments

When I was working in online video streaming, we were told we couldn't do CD because TV apps had to be certified by manufacturers. Yet we did CD. It helped us build enough trust so that some manufacturers allowed us to release most changes without certification, and others fast-tracked our certification process. When I was working on internal enterprise software, we were told we couldn't do CD because releases had to be planned and manually tested, deadlines respected, and users trained. Yet again, we did CD. Plans became high-level and abstract as iterative execution took over. Automated acceptance tests and feature flags allowed our QA testers to become our first users in production, rather than gatekeepers for releases. Training was built into the frontend applications. It may be the toughest challenge yet, but I'm determined to make it work in a business SaaS startup, where client demands and approvals are everything. The excuses are plentiful, but I have yet to see a domain where CD has failed to deliver results. After all, some of the most famous CD success stories are about printer firmware and an MMO game. Are you sure the software you are working on is more exotic than that?

19

I've been living in Claude Code for over a month now, and the results are... mathematically aggressive. My throughput has increased by roughly 5x. I'm shipping features in the three-minute gaps between meetings and running parallel workstreams like a distributed system that's slightly over-provisioned. On paper, I'm a high-performance machine. In reality, I'm exhausted. Here is the irony of the "AI-accelerated" era: the AI does the tactical heavy lifting, but I'm the one carrying the cognitive load. The Reality Check: Context Switching is the new CPU Spike: I'm not typing as much, but I am pivoting between five different architectural contexts a minute. My brain wasn't designed for this many concurrent threads. The PR Tsunami: Everyone on the team is now a high-velocity output factory. We are drowning in pull requests and updates. The Vigilance Tax: When code is generated at this speed, "attention to detail" stops being a soft skill and becomes a survival mechanism. We're reviewing more code with less oversight, which is a recipe for high-speed technical debt if we blink. The Looming Question If we're all shipping 5x more code, who is actually thinking 5x more about the long-term implications? I'm worried that we're trading physical keystrokes for mental burnout. We've automated the "doing," but we've accidentally hyper-scaled the "deciding." I don't have the answer yet, but "work harder" isn't the solution when the machine already outworks you. We need a new framework for managing mental bandwidth before we all hit a total system failure. How are you keeping up?

185

20 Comments

Being 10x faster doesn’t matter if you’re still producing zero leverage. Speed isn’t the bottleneck. Great article (first comment) that outlines why most people are optimizing tools while a few are redesigning how value gets created.

6

1 Comment