Software Engineering & Architecture Andrej Karpathy's talk at Sequoia's AI Ascent drew a sharp line between "vibe coding" and what he calls "agentic engineering" — designing systems with constraints and supervising AI agents as capable but fallible interns. His central argument: as agents produce larger and more coherent code chunks, human architectural judgment becomes more critical, not less, because design flaws now scale with agent productivity. The core principle: "Understanding cannot be outsourced, only thinking can." From Vibe Coding to Agentic Engineering
Global Macro & Markets The US and China announced a 90-day tariff reduction deal taking effect May 14: US tariffs on Chinese imports drop from 145% to 30%, and Chinese tariffs on US goods fall from 125% to 10%. Markets staged their biggest single-day rally in over a month — S&P 500 +3.2%, Nasdaq +4.3%, Dow +2.8% — compounded by separate reports that the US and Iran were nearing a deal that sent crude oil plunging ~8% intraday. President Trump cautioned both de-escalations could reverse if upcoming talks stall, with a Trump-Xi summit scheduled for May 14–15 as the next catalyst. S&P 500 six-week winning streak — CNBC
Enterprise AI & AI Strategy OpenAI and Anthropic both announced separate joint ventures backed by major private equity firms — Blackstone, Goldman Sachs, and others — totaling roughly $5.5B in combined capital, with both already in talks to acquire established AI services and consulting firms. The move signals that both companies see deployment expertise, not model capability, as the new bottleneck to enterprise AI ROI. It's a direct competitive strike at McKinsey, Accenture, and the broader $300B management consulting industry. Anthropic and OpenAI launch enterprise AI consulting joint ventures
AI Tools & Developer Utilities Anthropic shipped three new capabilities for Claude Managed Agents: "Dreaming" (research preview), which lets agents review past sessions and self-improve over time; "Outcomes" (public beta), where a separate grader evaluates task success against a developer-defined rubric reporting up to 10-point accuracy gains; and "Multiagent Orchestration" (public beta), enabling a lead agent to delegate to specialist sub-agents in parallel. Together these shift Claude agents from stateless per-session tools toward systems that learn, self-assess, and coordinate autonomously. MCP separately crossed 97 million installs and moved under Linux Foundation governance, cementing its role as the standard agent-tool connectivity layer. New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration