🎯 Top 3 Things to Know
1. The Model Context Protocol published a release candidate for its 2026-07-28 spec, the largest rewrite since the protocol launched. The headline change is that MCP becomes stateless at the protocol layer. Today a session pins a client to one specific server instance, which complicates load balancing and failover. The new spec drops that binding, so any server instance can answer any request. It also adds a formal extensions framework with reverse-DNS identifiers, and ships MCP Apps, an extension that lets a server render HTML inside a sandboxed iframe on the host. This matters for anyone running MCP servers in production, where the sticky-session model has been an operational headache. The maintainers set a ten-week validation window before the spec finalizes. Worth pulling the release candidate now and testing whether your servers survive requests arriving at any instance. MCP blog
2. OpenAI began serving GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second. Two days after GPT-5.6 reached general availability, the speed story arrived. Sol is OpenAI's frontier coding and agentic model, and 750 tokens per second is fast enough to change how agentic loops feel: less waiting between tool calls, quicker multi-step chains. The constraint this addresses is latency, which has quietly capped how many reasoning steps an interactive agent can take before a user gives up. Relevant to teams building agents where round-trip time, not raw quality, is the bottleneck. Worth timing a real agent workflow on the faster endpoint against your current one before assuming the quality tier alone decides the deployment. OpenAI news
3. The EU Council gave final approval to the Digital Omnibus, pushing back the AI Act's high-risk deadlines and adding new prohibitions. The simplification package cleared its last hurdle on June 29 and is expected in the Official Journal this month, ahead of the August 2 date when high-risk rules would otherwise bite. Stand-alone high-risk systems now have until December 2, 2027, and product-embedded ones until August 2, 2028. The package also bans "nudifier" tools that generate non-consensual intimate images or CSAM, effective December this year. It affects any company placing AI systems on the EU market, which gains time on high-risk compliance but faces a hard new content prohibition. Worth checking whether any deployed system touches image generation of real people, since that carries a December deadline regardless of the high-risk relief. Council of the EU
🚀 Frontier Models & Features
OpenAI shipped GPT-Live, a real-time voice mode that listens, thinks, and speaks at the same time rather than waiting for a full pause. It can interject short acknowledgments and handle quick back-and-forth, which moves voice agents closer to natural turn-taking. OpenAI news
Google DeepMind confirmed Gemini 3.5 Pro slips to July 17. The company scrapped the 2.5 Pro architecture for a full rebuild, citing math reasoning, SVG scene generation, and image quality. A delay for an architecture change, rather than a quiet slip, is worth noting. BigGo
🔬 Research Worth Reading
- Why Reasoning Fails to Plan: A Planning-Centric Analysis of Long-Horizon Decision Making in LLM Agents (Zehong Wang et al. / University of Notre Dame). arXiv
- TL;DR: Argues that step-by-step reasoning quietly turns into a greedy policy: the model scores each next step locally, commits early, and cannot recover when a later consequence punishes the early choice. The fix, FLARE, forces explicit lookahead and value propagation so downstream outcomes influence early decisions.
- Stat: With FLARE, LLaMA-8B frequently outperforms GPT-4o using standard step-by-step reasoning on long-horizon planning benchmarks.
- Apply it: If an agent handles multi-step tasks well in isolation but derails on long horizons, add a lookahead step that estimates the value of a few candidate action sequences before committing to the first move, and measure recovery rate on tasks with delayed consequences.
🏢 Enterprise in the Wild
Quiet day on this front. Survey-level 2026 data continues to circulate (Gartner projecting embedded agents in 40% of enterprise apps by year end), but no fresh single-company production case study surfaced with traceable outcomes.
🛠️ Tooling & Ecosystem
AgentPrizm launched AgentMemory and AgentSkills, a platform pairing a REST API with MCP infrastructure to give agents persistent memory across sessions. Persistent memory remains one of the weaker links in production agent stacks, so a dedicated layer for it is worth a look. Agentic.ai
M-Files shipped M-Files for Consulting, a document-governance app that tracks consulting work from proposal to delivery and gates access to confidential client material. Agentic.ai
⚖️ Policy & Regulation
Beyond the Digital Omnibus above, the US backdrop remains a push toward preemption: a December 2025 executive order directed agencies to challenge state AI laws seen as inconsistent with a uniform national policy, and a March 2026 White House framework recommended Congress legislate broad preemption under a light-touch standard. The transatlantic gap between the EU's rule-based regime and a lighter US posture continues to widen. Collibra
📌 Watch List
- Stateless agent protocols: MCP dropping session affinity reshapes how servers scale.
- Real-time voice agents: simultaneous listen-think-speak moves past strict turn-taking.
- Long-horizon planning: lookahead methods challenge pure step-by-step reasoning.
- Persistent agent memory: dedicated cross-session memory layers gaining traction.
- Mid-tier cost models: Terra and Luna extend the cheaper-but-capable frontier.