MMARW / INTELLIGENCE / AI
Analysis of artificial intelligence, models, agents, products, and the systems reshaping how work gets done.
AI · Oct 9, 2026 · 10 min read
Broadcom's up-to-$42bn loan to Anthropic ties chip supply to model supply. How firms built on frontier-model APIs can manage the vendor risk.
AI · Oct 7, 2026 · 8 min read
Build an MCP tool registry that lets Ops approve agent tools fast: ownership, risk tiers, and a path that does not freeze delivery.
AI · Oct 7, 2026 · 9 min read
Design agent permission scopes that pass client audits: least privilege, clear boundaries, and reviewable access without slowing delivery.
AI · Oct 6, 2026 · 5 min read
After Q3 2026 CapEx noise, the defensible FY2027 agent budget funds proof of safe completed work — not a bet that hyperscaler spend makes agents cheap.
AI · Oct 5, 2026 · 6 min read
Hyperscaler CapEx is a leading signal for AI-services buyers. This file shows how to read Microsoft, Alphabet, and Meta spend for capacity, scarcity claims, and negotiation evidence — without treating companywide CapEx as rentable GPUs.
AI · Sep 28, 2026 · 5 min read
A client demo that survives is repeatable artifact production plus failure handling — not a lucky click path. Choose Computer Use vs browser agents by surface, then scope stop rules, human gates, and the sandbox.
AI · Sep 21, 2026 · 2 min read
Treat model-named sources as claims, require openable URLs for load-bearing cites, stamp as-of dates, and run a short pre-ship citation audit before you send.
AI · Sep 14, 2026 · 4 min read
Competitor research dies in the browser. This piece shows how to replace a 40-tab pile with a file: one question, bounded scope, comparison dimensions, openable sources, gaps, and a recommendation someone else can act on.
AI · Sep 3, 2026 · 5 min read
A chatbot answers the next message. An agent takes a job and tries to finish it. This piece is the practical split: when a chatbot is enough, when you actually need an agent, and what breaks if you pick wrong.
AI · Sep 2, 2026 · 6 min read
Long-horizon agents fail when intent evaporates between calls. This piece covers how to keep a plan stable, distinguish consumed edges from new work, and implement scratchpads, receipts, replay vs replan, and drift evaluation.
AI · Aug 31, 2026 · 3 min read
Chat is cheap. A saved document with a URL is the actual unit of work. This short Intelligence piece argues that agents only become useful when they produce inspectable artifacts, and what that means for teams still living in the thread.
AI · Aug 22, 2026 · 23 min read
AI agents are goal-directed systems that use AI to execute multi-step tasks through tool use and decision-making.