Writing on AI agent monitoring, developer productivity, and the ROI of AI coding tools.
Git shows you what landed. It has no record of what got thrown away, and that is where a large share of your AI spend lives.
Amodei, Altman, and Musk agree AI should slow down. For founders, the real 2026 question isn't p(doom) - it's whether your AI spend is making engineers measurably faster.
Over 15% of AI-authored commits introduce issues, yet lab studies show 55% speedups. Here's how to measure the real ROI of Copilot, Cursor, and Claude in 2026.
Measuring developer productivity is hard once AI writes the code, because the usual metrics break. Here is what actually shows whether your AI spend is paying off.
McKinsey's new Rewired excerpt maps a 4-level productivity curve topping out at 20x. Most teams are stuck at 1.2x - and can't measure it. Here's why observability, not tool adoption, separates the winners.
Most teams measure AI activity, not AI productivity. Learn why seat counts and prompt volume mislead leaders, and what to measure instead.
A practical framework for measuring AI developer productivity across Claude, Cursor, and Copilot - the metrics that correlate with shipped work, and the vanity numbers to ignore.
A grounded look at whether Claude, Cursor, and Copilot improve developer productivity - where the gains are real, where they're inflated, and how to tell which is which on your team.
A structured AI governance framework for engineering teams: how to monitor Claude, Cursor, and Copilot adoption, attribute spend, catch misuse, and prove ROI - without slowing engineers down.
The dashboards are lying to you. Not on purpose - they're just measuring the wrong thing. In an AI-leveled engineering org, looking like a 10x engineer is no longer evidence of being one.
Y Combinator says every company needs a Company Brain. True, but incomplete. A static skills file is a guess wearing a tie until you instrument the agent that consumes it. Five stress tests to run first.
METR's viral chart shows AI capability doubling every seven months. The real story is the 39-point gap between perceived and measured developer speedup - and why your agents need real performance tracking.
Only one-third of organizations have reached governance maturity on agentic AI. Here's what McKinsey's 2026 numbers mean for every enterprise AI procurement cycle.
Jensen Huang says your agents will harass and micromanage you. He's right - and the missing primitive is evaluation. Here's why AI agent monitoring is the real frontier.
GPT-5.4 scored 83% on GDPVal - a benchmark testing 44 real professional occupations. But a lab benchmark and a production deployment are two very different things.
Two of the most important AI companies sacrificed their flagship products and placed the same bet. The infrastructure to evaluate and monitor these agents barely exists.
You configured it, tested it, and scheduled it. But if someone asked you exactly what it did on its last run - could you answer that?
OpenClaw amassed 135,000 GitHub stars in weeks - but the real crisis isn't security. It's that nobody knew what their agents were actually doing.
Jensen Huang declared AGI is here on the Lex Fridman Podcast. But the real question isn't whether AGI has arrived - it's whether anyone can actually prove their AI agents are working.
You set up the agent. It runs. But is it actually doing what you need it to do? Most non-technical founders have no reliable way to answer that question. Here's how to change that.
Jensen Huang just told 320,000+ viewers that every engineer will soon run a fleet of AI agents. That's not just a productivity prediction - it's a monitoring crisis in slow motion.
Your AI agents are running 24/7. But do you actually know what they're doing? Most teams discover failures days after they happen.