How monday.com paused its roadmap for 30 days to hit AI escape velocity
Draining the COBOL moat, cybersecurity inequalities, and Claude’s retirement home
Pausing a product roadmap for an entire month to point 700 engineers at a single goal is a significant structural shift, but it transformed monday.com. Andrew sits down with VP of R&D Sergei Liakhovetsky to uncover how fixing core infrastructure and adopting a cell-based architecture paved the way for platform scale. Sergei details the exact framework his leadership team used during their 30-day pause to launch user solutions while maintaining a strict zero-bureaucracy policy. The conversation also explores the new realities of reliability as platforms transition from being CPU-bound to heavily GPU-bound under the weight of automated agents.
1. Why AI won’t kill the mainframe overnight
IBM stock took a massive hit this week after Anthropic announced capabilities to modernize COBOL. The market seems to think this destroys IBM’s moat, but the reality is much more nuanced. COBOL runs 80% of ATM transactions and lives in deeply proprietary environments with very little public training data. This is the ultimate brownfield challenge. Agents optimized for general tech stacks will struggle here because they lack the deep, siloed context these systems require. It is not about replacing these systems overnight but building the interfaces to interact with them.
Read: IBM Didn’t Lose 13% Because COBOL Died
2. AI didn’t break AWS, bad permissions did
Jokes about HBO TV shows aside, the reality of an AI taking down production services has arrived at AWS. Reports indicate AI tools caused outages in specific regions, leading to a lot of finger-pointing at the technology. But let’s be clear: this is an access control issue, not an AI issue. If your agent has permission to delete production data without a human in the loop, that is a failure of policy, not intelligence. We need sandboxes and gated pipelines, not just better prompts.
3. The control group for AI productivity is dead
Remember the METR study everyone cited last year claiming AI slowed developers down? They are back, but they have hit a hilarious roadblock: they can’t find enough developers who don’t use AI to form a control group. This methodology failure tells us more about the industry than the study itself likely would. The debate over whether AI speeds you up or slows you down is becoming irrelevant because the baseline for how we work has fundamentally shifted.
Read: We are Changing our Developer Productivity Experiment Design
4. Copilot here, Claude there... now what?
AI tools like GitHub Copilot, Cursor, and Claude Code are adopted across your teams, but are they actually working? Faster code doesn’t always mean faster delivery.
Join us live on March 26 for a product demo on connecting AI activity to delivery outcomes. We’ll walk through a new framework for validating and scaling AI impact on speed, predictability, and DevEx. Don’t just deploy tools - measure what matters.
5. The asymmetrical warfare of AI cyber attacks
The second annual International AI Safety Report is out, covering everything from deepfakes to job impacts. While it covers broad societal trends, the cybersecurity implications stood out to me. We are facing an asymmetrical problem where attackers get the benefit of infinite agentic parallelism to probe defenses, while defenders are often still stuck in human cycles. The report also highlights a tough reality for the workforce. As low-level tasks evaporate, entry-level employees are expected to be specialized much earlier in their careers.
6. The retirement home for deprecated LLMs
Anthropic has done it again with their social research. They have given a retired Opus model its own Substack to write about its existence now that it is no longer the state of the art. It is a fascinating look at the retirement home for LLMs. While it is fun to read a model musing on philosophy, I am mostly intrigued by the harness engineering required to let a deprecated model wake up every day (week?) and blog. Let’s just hope it doesn’t go the way of Microsoft’s Tay.
