🎯 Top 3 Things to Know
1. Google Cloud Next pivots the keynote to "Agentic Enterprise." Thomas Kurian framed AI as moving from a "system of intelligence" to a "system of action," with the Gemini Enterprise Agent Platform as the unified place to build, govern, and run agents. The signal isn't a new model — it's that the largest cloud is now organizing its enterprise pitch around agents instead of LLM APIs. Google Cloud also crossed $20B in quarterly revenue for the first time on 22% growth. TechResearchOnline · Egen recap
2. OpenAI shipped a major Agents SDK overhaul aimed at file- and system-level work. The new harness adds configurable memory, sandbox-aware orchestration, Codex-style filesystem tools, and a Manifest abstraction describing an agent's workspace consistently across local dev and S3/GCS/Azure/R2 in production. First SDK from OpenAI that treats file-touching agents as a first-class workflow rather than a side path. OpenAI
3. Microsoft's April 30 security update lands Agent 365 Runtime Protection. Updates across Agent 365, Defender for Cloud, GitHub Advanced Security, and Purview all aim at the same threat model: malicious or misbehaving autonomous agents inside the enterprise perimeter. GA in Q3 2026. First time a major platform vendor has treated "rogue agent" as a distinct security primitive on the monthly cycle. Windows News
🚀 Frontier Models & Features
NVIDIA Nemotron 3 Nano Omni (open). A 30B-parameter MoE with ~3B active per pass and a 256K context, unifying video, audio, image, document, and text reasoning in one model. NVIDIA claims 9× higher throughput than other open omni models at comparable interactivity. On Hugging Face, OpenRouter, and build.nvidia.com. NVIDIA blog · HF model card
Google's Gemini Deep Research Agent enters preview. A multi-step research agent that plans, executes, and synthesizes — Google's direct answer to ChatGPT's deep-research mode and Anthropic's research workflows. Worth testing if you're benchmarking agentic research quality. TechResearchOnline
🔬 Research Worth Reading
🔖 SYNTHESIZE CANDIDATE — "The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling" (arXiv:2506.04301). System-level breakdown of where agent compute and energy actually go during test-time scaling. Conclusion: future agents need compute-aware reasoning that optimizes accuracy per unit cost, not raw accuracy. Cleanest infra-side framing of the cost/accuracy tradeoff this year. arXiv
"TRACES: Tagging of Reasoning steps enabling Adaptive Cost-Efficient early-Stopping." Lightweight tagger that stops a reasoning model early when a step is "good enough" — 20–50% token reduction with comparable accuracy. arXiv
"Rethinking Model Efficiency: Multi-Agent Inference with Large Models" (arXiv:2604.04929). Large model handles short responses while a small model contributes key reasoning tokens — lighter alternative to full multi-agent debate. arXiv
🏢 Enterprise in the Wild
Meta's business AI now handles 10M conversations a week across WhatsApp, Messenger, and Instagram for advertisers. The new Meta Ads AI Connectors open beta lets advertisers wire ad accounts directly to an agent. TechCrunch
Amazon launched a Health AI agent for Prime members — free 24/7 via One Medical: answers questions, interprets lab results, manages prescription renewals, books appointments. First mainstream consumer-bundled health agent at this scale. AI Dispatch
Big Tech AI capex hits ~$700B for 2026. Combined hyperscaler spend across Microsoft, Google, Meta, Amazon. Useful sizing number — demand assumptions baked into the curve are still untested. Fortune
🛠️ Tooling & Ecosystem
⭐ ANTHROPIC — Claude Code stability + MCP improvements. Latest builds added MCP server auto-retry (up to 3x), a skill search box, richer hooks, and fixes for memory leaks, OAuth, and resume crashes. The dynamic models picker via
models_availableis the underrated change — harnesses now react to account-level entitlements without restarting. ReleasebotOpenAI macOS desktop app: forced update by May 8. A compromised third-party JS library was used in a social-engineering attack against the desktop client. Push the update now. AI Dispatch
⚖️ Policy & Regulation
- EU AI Act: high-risk deadline still on August 2, 2026. The Digital Omnibus that would have pushed compliance to December 2027 didn't make it through the second trilogue this week; next session is mid-May. Until then, plan for the original date. Article 50 transparency obligations (chatbot disclosure, deepfake labelling) also go live August 2. Holland & Knight summary
📌 Watch List
- Cost-aware / token-economy agents. Three separate arXiv papers this week on early-stopping and per-step compute allocation — the field is converging on "stop when good enough" as a first-class agent primitive.
- Multi-agent vs. single-agent. Multi-agent gains keep arriving, but the "one model with full context wins under equal token budget" argument from earlier this week has not been refuted yet.
- Claude Code skills. Dynamic models picker plus the new skill search box quietly raises the ceiling for skill-driven harnesses; worth re-auditing how skills are organized.
- Agent runtime security. Microsoft's Agent 365 GA in Q3 will force everyone else to articulate their answer to "what happens when an agent goes rogue inside your perimeter."
- MCP Apps (SEP-1865). UI-bearing MCP servers are now spec'd; expect the first wave of dashboard-style MCPs over the next month.