Short, sober pieces from regulated-AI delivery — the failure modes, the controls that fix them, and the regulation closing in. No hype, no "future of AI." Evidence over adjectives.
CB Insights mapped 400+ AI agent startups. Look closer and you see the beginning of a collapse — a gravitational pull toward vertical specialization. Why depth is overtaking breadth, and why the territory is collapsing into something more real.
Read the field noteThe frontier stopped prompting and started building loops. They're right — a loop is a control plane. Here's the governance it still needs: a verifier that proves its claim (AUD-1), a kill threshold before the bill (LMT-1), and a human who stays in the decision (HITL-3).
The 2:1 paradox, the healthcare gold rush, the M&A consolidation wave, and the three engineering pillars that separate durable companies from acquisition targets. Depth is overtaking breadth.
A pattern that destroys careers and burns runway: a company hires a “Head of Growth,” both sides celebrate, and 12–18 months later it unravels. What actually goes wrong, and how to structure the role so it doesn't.
Anthropic's practical blueprint for real-world agents, distilled — from memory design to orchestration. The difference between a demo and a production system.
While the feeds filled with “new year, new opportunity” posts, Meta quietly acquired an AI agent startup processing 147 trillion tokens. The window is open; here's how not to squander it.
Most enterprise AI pilots fail before production. The models work. The technology is impressive. What's missing is governance, audit trails and kill switches — and that's a build problem, not a research one.
Why prompt-level safety fails under a regulator — and how least-privilege service accounts block destructive actions at the keyring instead of asking the agent nicely. The Najm Insurance pattern.
The read-only ledger that makes every agent decision traceable and reversible in under 30 seconds — the audit trail built to survive a national government, not satisfy a dashboard. From the NMML estate.
High-stakes actions pause for a human; everything else runs free. How De Lijn got a board-approved roadmap across 5,000+ FTE without exposing leadership to irreversible autonomous decisions.
Continuous telemetry on speed, spend and error volume, with automatic suspension on breach. How a LangGraph multi-agent system replaced 200 FTE with agents that can't cross a boundary they don't know exists.
EU AI Act high-risk obligations — risk management, logging, human oversight, transparency — and a practitioner's read on preparing for them while the Digital Omnibus revisions are still pending.
On these notes: pieces are drawn from real regulated engagements; client specifics are generalised and figures match the published case studies. The longer essays live indexed in the Knowledge Hub.