posts · 2026-08-21

This piece lays out the awkward math behind big tech's agent strategies. Microsoft's GitHub Copilot and M365 Copilot share similar tech foundations, but the outcomes diverge sharply: 20M users vs. 3.3% paid penetration and 20–30% weekly active usage. The split isn't about capability—it's about the revenue model. GitHub's agent drives more commits and CI minutes, so usage revenue grows with output. An M365 agent that automates reports and invoice checks would let enterprises cut E5 seats at $57/user/month, so Microsoft's price sheet only ever offers per-seat add-ons, never outcome-based pricing. Google shut down its standalone browser agent Project Mariner and folded the pieces into Search and Chrome—an agent that does price comparison and checkout means fewer search result page views and ad clicks. Meta put revenue-generating Business Agents on ad-free WhatsApp; personal agents remain free tests because they'd cannibalize feed ad exposure. Amazon has the best incentive alignment—agent-driven purchases boost GMV and fulfillment revenue—but Alexa+ was delayed two years and saw low voluntary use even after a forced rollout. The litmus test to remember: the better this agent works, does the company make more money, or less? If the answer is less, it'll only ever be sold as a surtax on existing subscriptions.

This essay earns its read by splitting tech-world 'impressive' into two clean categories and walking through four case histories that make the distinction stick. ChatGPT launched in the most hated product form—a chatbot—with a mechanism so boring (next-token prediction) that no one could compress its value into a sentence. It pulled 100M users through 30-second self-demos. AutoGPT did the reverse: one grand sentence about autonomous task completion, 100K GitHub stars, and a codebase that was basically a for-loop around an API call, with failure compounding every iteration. GraphRAG looked brilliant on both fronts—the sentence about restoring relationships flattened by embeddings, the mechanism of entity extraction and community detection—but indexing a small text chunk could burn $15, and an independent eval across 20 datasets showed a 67 vs. 69 split against baseline RAG. MCP's 'USB-C moment' pointed at elegant standardization; the real value that shipped was crude but functional tool distribution through a one-line config. The author offers three mechanisms driving the divergence: selection effects (products that let users verify value instantly don't compete on slogans), condition-stripping (a sentence goes viral by deleting the exact constraints that predict whether it'll work for you), and expectation asymmetry (grand claims turn minor delivery gaps into failures—IBM Watson Health burned $62M at MD Anderson without treating a single patient, while Databricks' boring RAG definition became the backbone for 60% of orgs). I'd use this as a mental filter: next time a tech claim gives you chills, ask whether you can verify it yourself in 30 seconds. If not, the sentence's viral power probably came from the constraints it deleted.