🎯 Top 3 Things to Know
Google ships Gemma 4 — open weights, native multimodal, four sizes. The drop on May 4 was the headline open-source release of the week: a family spanning E2B/E4B (edge), 26B MoE, and 31B dense, with day-one support across vLLM, llama.cpp, MLX, Ollama, and Hugging Face. The 31B sits at #3 on the Arena open leaderboard. All sizes natively handle video and images, edge variants take audio in, and the larger models extend to 256K context. (Google blog)
Compute-normalized study: single agents match or beat multi-agent systems on multi-hop reasoning. Provocative new arXiv paper argues that recently reported multi-agent gains are mostly an artifact of extra test-time compute. When thinking-token budgets are equalized, a single agent does as well or better. If it holds up, this changes the default architecture choice for a lot of harnesses. (arXiv 2604.02460)
EU AI Act high-risk obligations on track to slip from August 2026 to late 2027. The Commission's November proposal to push the high-risk timeline by up to 16 months is moving forward, citing missing standards and implementation tooling. Companies preparing high-risk conformity work should reset planning assumptions; prohibited-practice rules and GPAI obligations remain in force. (European Commission AI Act page)
🚀 Frontier Models & Features
Gemma 4 (Google). See Top 3. Notable for the agent stack: native function-calling, structured JSON output, system instructions baked in, Apache 2.0 license. (blog)
Anthropic Claude Opus 4.7 — document reasoning graph update (May 4). Quiet incremental improvement to how Opus 4.7 handles multi-document reasoning; pricing unchanged ($5/$25 per M tokens). The new tokenizer continues to push token counts 12–27% higher than 4.6 on long inputs — worth re-checking your cost models if you migrated. (release notes)
🔬 Research Worth Reading
SciResearcher: Scaling Deep Research Agents for Frontier Scientific Reasoning (Tsinghua + collaborators). Fully automated agent framework for synthesizing conceptual + computational science tasks; SciResearcher-8B hits 19.46% on HLE-Bio/Chem-Gold, SOTA at its scale. Useful read if you're building research agents that need to reason over actual papers. Verdict: read full paper. (arXiv 2605.01489)
Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets. Normalizes for compute and finds the multi-agent advantage largely vanishes. Direct relevance to anyone defaulting to multi-agent for complex reasoning. Verdict: 🔖 SYNTHESIZE CANDIDATE — it punctures a common assumption with a clean experimental setup. (arXiv 2604.02460)
Are Latent Reasoning Models Easily Interpretable? Probes whether continuous latent reasoning is a black box; finds current latent-reasoning models still encode mostly interpretable processes, and interpretability itself correlates with prediction correctness. Verdict: read abstract. (arXiv 2604.04902)
From Actions to Understanding: Conformal Interpretability of Temporal Concepts in LLM Agents. Step-wise conformal framework for labeling agent internal states as on-track or failing. Could plug into harness instrumentation. Verdict: skim. (arXiv 2604.19775)
🏢 Enterprise in the Wild
Ambient AI clinical documentation in production. A health system rolled out real-time transcription and structuring of clinical notes inside the existing EHR, targeting the 2–3 hours/day physicians spend on documentation. Reported timeline: six months end-to-end, with eight weeks of structured change management before pilot. The cited stat from the broader study: skipping change management correlated with 40% lower adoption. (Stanford Enterprise AI Playbook)
Financial services agent for meeting follow-through. Same study: a financial services firm deployed agentic workflows that pull commitments from video conferences, draft reminder communications, and track follow-through. Notable as a concrete, narrow agent scope rather than open-ended copilot. (same source)
🛠️ Tooling & Ecosystem
⭐ ANTHROPIC — Claude Agent SDK for Python (May 4). Released with parallel MCP server reconnection (replaces serial), PostToolUse hooks that can rewrite tool output for any tool (was MCP-only), duration_ms on hook inputs for execution timing, and a fix for the SDK hang on malformed parallel tool calls. (Claude Code changelog)
Agent harness deep-dive (Upadhyay, May 2). Walks through building an agent harness from scratch — control plane, tool registry, verification, retry logic. Lift for your own harness: the explicit "every mistake the agent makes becomes a hard guard" pattern (Hashimoto-style) plus a clean separation between tool-routing layer and verification layer. (atalupadhyay.wordpress.com)
⚖️ Policy & Regulation
NIST publishes final Adversarial Machine Learning report. Voluntary guidance with a taxonomy of attacks on AI systems — useful baseline for any team writing an AI risk register or threat model. (Mintz writeup with link to NIST)
EU AI Act high-risk delay (see Top 3). (European Commission)
📌 Watch List
- Cost-aware agent design — dynamic turn limits cut cost ~24% at equal solve rate; flexible thinking budgets are now mainstream in research. (arXiv 2602.18998)
- Multi-agent vs single-agent debate is suddenly live again — compute-normalized comparisons are the new standard. (arXiv 2604.02460)
- World models / JEPA — V-JEPA 2 and ThinkJEPA continue the thread of latent-world-model + VLM hybrids. (arXiv 2506.09985 · arXiv 2603.22281)
- Harness engineering as a discipline — "every mistake becomes a guard" is consolidating into named practice. (OpenAI on harness engineering)
- KV-cache compression — Google's TurboQuant (ICLR 2026) targets the long-context memory bill. (No paper yet — tracked via conference program.)
- Interpretability of latent reasoning — early signal that "latent ≠ opaque." (arXiv 2604.04902)