You’re reading the first edition of our monthly news article that serves up the most substantial, thoughtful, and useful AI content – with a healthy dose of skepticism. No hype or hand-waving (even though I do enjoy waving my hands), no chasing every new model, no “top 50 tools” lists that will die faster than you can open a JIRA ticket to implement them. Only things that can genuinely level you up as engineers, leaders, and tech-immersed humans. Think slow-food for your FOMO.
Man, I’m getting older, and I started wondering what the world needs so that my four-year-old daughter’s generation has it as good as possible when they grow up. This thought experiment led me to the same conclusion as the rest of LinkedIn: the world needs another AI news article series.
I’ve decided to take on the challenge. Not every hero wears a cape.
Those who know me know I’m a walking synonym for FOMO. Every morning, before 6:00 a.m., I scan everything that’s happened on the internet – from Hacker News and a few newsletters to other sites. It’s my morning routine, before I drop my daughter off at preschool. Thanks to that, I fish out gems that rarely hit the mainstream, yet make for interesting case studies or technical inspiration. I like sharing them on the company Slack, and now I’ll share them on the company blog, too.
I realize the space is already crowded, and I don’t intend to race with experts who label everything “groundbreaking” the minute benchmarks drop. So I’ve set a few simple rules:
No hot “bleeding edge” news New OpenAI models or open-source projects from China will reach you in a flash anyway. And extreme experiments are rarely practical.
If white papers, then only survey-style I value white papers and deeply technical reports, but for most of us they’re too detailed. I’ll favor surveys and articles you can quickly “chew on” and put to work day-to-day.
Few tools If a toolkit appears, it’ll be truly valuable and vetted– - not a list of everything that moves. In my experience, few novelties survive effective market triage, and their half-life is roughly the interval between issues of a weekly newsletter.
Mostly long-form pieces – readable, practical deep dives on topics useful to engineers. I prefer thorough, nuanced articles to short news bites, because only the former cut through the noise and let you understand a topic beyond the basics. From experience, writing a few pages of genuinely valuable text takes effort– - I hope that effort turns into real benefits for you.
I’ll also focus on topics that are actually useful for Senior Engineering Managers and Tech Executives. So expect discussions of trends, key architectural patterns, and lessons learned from practice. We’ll sometimes veer off that path, but that’s the audience I have in mind. If that’s you, I hope you’ll enjoy July’s best picks.
Oh, and the whole thing will, of course, be unapologetically opinionated – like all good things should be. 🤷
Let’s start with a healthy dose of contrarianism– - even though every 2025 headline promises a revolution in autonomous AI agents, Utkarsh Kanwat, in his piece Why I'm Betting Against AI Agents in 2025 (Despite Building Them), bluntly argues that this narrative doesn’t match reality.
The limits of autonomy and economics
He explains that compounding errors in multi-step processes (95% reliability per step yields only 36% success after twenty steps) and the quadratic token costs of “chatty” agents make mass deployments economically and mathematically infeasible. Instead of chasing every new model, he shows how constrained-scope systems (3–5 steps, checkpoints, rollbacks) and stateless “function generators” can genuinely save hours of work without bankrupting companies.
Engineering – not magic
Kanwat also dissects the other pillar of the hype –-- tooling and integration with messy enterprise ecosystems. He stresses the importance of feedback loops between the API and the agent, tool-state management, communicating complex operations without clogging context, and robust recovery mechanisms. His database agent illustrates this well: it isn’t a “single-shot” question, but a multilayered system of transactions, rollbacks, query queuing, and audit—where the AI drafts the query, but the real magic comes from traditional systems engineering.
Practical advice for builders
Finally, he offers advice for those who’ll build agents anyway: avoid boundless “full autonomy,” set clear contracts between AI and humans, design for failures as carefully as for successes, and always account for token costs. He predicts startups promising impossible 20-step workflows will fold, while winners will focus on narrowly defined, economical, and reliable assistants that act as a communication layer atop classic engineering.
And the last piece of advice here is: don’t bite off more than you can chew—and don’t forget that software development (and an agent is software) is an iterative process.
The three layers of ROI
In the article The Three Layers of ROI for AI Agents, Henry Pray shows that while AI agents immediately reduce the cost of repetitive tasks, the real savings appear only over time– - once the team gets used to the new way of working and process improvements are actually adopted and measured.
At the same time, agents open up entirely new revenue streams by taking on tasks that previously weren’t worth delegating—from cold sales outreach, through automatic proposal and RFP generation, all the way to reactivating dormant leads.
Deeper returns come with iterative optimization.
Iteration is the real superpower
Once the agent system is embedded in day-to-day work, it’s worth focusing on improving the entire workflow– - from scenario simulation and predicting bottlenecks to continuous refinement based on hard data like customer satisfaction and response time.
Through disciplined prototyping, testing, and measuring second-order effects, Henry Pray shows that ROI really takes off only when each subsequent improvement results from a deliberate cycle of learning and refinement, not a one-off rollout. In other words, the good old rule from Master Tolkien…
…at least as long as your runway hasn’t run out (which also fits with Tolkien canon).
But so no one mistakes me for a caveman running around with a club while everyone else is flying spaceships—I’ve actually grown quite fond of Vibe Engineering (since I can’t really claim to be doing capital-C Coding). That’s why my next two pieces will be on that very topic.
Creating with guardrails
In How to Vibe Code as a Senior Engineer, Alex MacCaw reveals the magic of AI-generated programming which, in the hands of an experienced senior, turns the drudgery of writing hundreds of lines of code into real fun—something many seniors have missed for years. I co-sign that wholeheartedly.
But the true power of “vibe coding” only shows up once you set solid guardrails.
The hybrid model of coding
The most valuable lesson from Chan is learning to explore the design space—and knowing when AI can lead and when it needs to step aside for the engineer. Code performance, error handling (race conditions, bloom filter setup, transactions), the cost of experiments, and “matters of taste” all still require human judgment, even with 97% automation.
His case study shows that agents are best assigned the repetitive parts, while critical checkpoints always get manual approval. It’s not fear of AI– - it’s the healthy hybrid of tool and engineer that leads to building truly modern systems.
But engineering is only part of it; what really matters is that someone has to use these AI systems and tools– - and that someone is still a human… at least for now.
That doesn’t mean the world of programming won’t change. In The software engineering "squeeze" Anton Zaides reminds us that for years being a software engineer was like a life hack—you could study intensively for a year and jump into the top 10% of earners.
Today, however, the “fat middle” of the market is already crowded with average developers (not you, dear reader), whose work will soon be marginalized by automation and rising employer expectations.
Stories of engineers losing roles to AI or bouncing off hundreds of rejections aren’t clickbait, but the real effect of oversupply and a rising cost of capital– forcing companies to optimize staffing and personnel budgets, and signaling to employees that they’ll be expected to support the company’s finances in different ways than before.
New players enter the game
Just two weeks later, Perplexity unveiled Comet – its answer to Dia – and OpenAI will soon join the fray. Comet leans on a default, built-in AI search that focuses on concise result summaries, and it offers a “Comet Assistant” in a side panel that can analyze emails, calendars, or on-page interactions. Although the two projects take different approaches – Dia as a native browser with AI in the background, Comet as an extension of core search – it’s clear we’re heading into an era where standard tabs and links give way to intelligent web assistants, and the future of browsing will be as thrilling as it is unpredictable.
I can’t wait either. It’s wild how little AI capability mainstream browsers have – and that it’s easier to paste text into a chat box than to get native summarization or a quick operation on highlighted text. We’re living like animals here in Chrome 🙈
Though maybe that’s still better than what Safari has to offer.
See you next month 😁
What top engineers still offer
Even so, Zaides stays optimistic– - he argues the market still needs true masters (that’s you, dear reader), not “code monkeys.” Those of us who can stand behind our code and also know the product, lead projects, and have a sense for design will be more in demand than ever.
Ambitious engineers who don’t wait for detailed tickets but define problems themselves, experiment with AI, and solve real challenges will pull ahead. And while the era of months-long coding bootcamps is probably over, the doors are still wide open for those who truly want in.
The rise of AI browsers----
We weren’t supposed to talk tools, but there’s a trend I’m following with real excitement, and I’m curious where it’ll take us– - AI browsers.
Vibe-enhanced browsing
M.G. Siegler of Spyglass wrote in Begun, the AI Browser Wars Have that he ditched his beloved Arc for Dia, the first fully AI-centric browser from The Browser Company. Although he still misses “Little Arc,” he quickly appreciated that Dia knows the context of all open tabs and can intelligently summarize their contents, eliminating copy-paste in day-to-day work.
That seamless AI integration into the UI makes work faster and more pleasant– - and this is only the beginning of a war for the next generation of browsers, where UX and contextual intelligence will be the keys to victory.
New players enter the game
Just two weeks later, Perplexity unveiled Comet – its answer to Dia – and OpenAI will soon join the fray. Comet leans on a default, built-in AI search that focuses on concise result summaries, and it offers a “Comet Assistant” in a side panel that can analyze emails, calendars, or on-page interactions. Although the two projects take different approaches – Dia as a native browser with AI in the background, Comet as an extension of core search – it’s clear we’re heading into an era where standard tabs and links give way to intelligent web assistants, and the future of browsing will be as thrilling as it is unpredictable.
I can’t wait either. It’s wild how little AI capability mainstream browsers have – and that it’s easier to paste text into a chat box than to get native summarization or a quick operation on highlighted text. We’re living like animals here in Chrome 🙈
Though maybe that’s still better than what Safari has to offer.
