🎯 Top 3 Things to Know
1. Google shipped Gemini 3.6 Flash, its cheaper workhorse model, cutting output prices while raising benchmark scores. The new Flash uses 17% fewer output tokens than 3.5 Flash on the same work, and Google dropped output pricing from $9.00 to $7.50 per million tokens, holding input at $1.50. Scores moved too: 49% on the DeepSWE coding benchmark, up from 37%, and 83.0% on OSWorld-Verified computer use, up from 78.4%. The knowledge cutoff advances to March 2026. This matters most for anyone running high-volume inference, where token count is the bill. Worth re-running your cost-per-quality numbers on Flash before committing to a pricier tier. Google also released a Flash-Lite and a security-tuned Flash Cyber, and said pretraining for Gemini 4 has begun. 9to5Google
2. OpenAI launched Presence, a managed platform for running production voice and chat agents. Presence changes how OpenAI sells to enterprises. Instead of raw model access, it bundles policies, guardrails, escalation rules, simulations, and evaluations into one governed system. Each deployment is scoped to a single job, such as billing or IT tickets, and given only the data and access that job needs. OpenAI is running it on its own English-language phone support line, where it says the agent resolves 75% of inbound issues without a human. It ships as limited general availability through forward-deployed engineers and integrators, not self-service. Teams weighing whether to build agent governance in-house should study what Presence bundles before rebuilding it. OpenAI
3. The Model Context Protocol published a release candidate that makes the protocol stateless. MCP is the standard for wiring models to tools and data, and the next spec, final on July 28, drops the requirement for sticky sessions. A remote MCP server that once needed a shared session store and gateway inspection can now sit behind a plain round-robin load balancer. The candidate also adds an Extensions framework, Tasks, MCP Apps, and hardened authorization. For anyone operating MCP servers at scale, this removes a real source of infrastructure pain. Read the candidate now and check whether your session-handling code can be simplified before the final ships. MCP blog
🚀 Frontier Models & Features
- Anthropic's Claude Developer Platform expanded Managed Agents: per-request model effort settings, wider webhook coverage for environment and memory-store events, session seeding with initial events, optional version checks on updates, and event deltas for thread streams. Anthropic release notes
- Google's security-tuned Gemini 3.5 Flash Cyber, aimed at finding software vulnerabilities, is being restricted to governments and trusted partners rather than shipped broadly. GCN
🔬 Research Worth Reading
Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task Allocation (Zhou, Leonardis, Feng / University of Birmingham). arXiv
- TL;DR: Instead of a fixed router deciding which model or tool handles each reasoning step, candidate agents bid for steps in an auction based on their own rectified confidence, and the highest bidder takes the task.
- Stat: Reports improvements over baseline routing across five benchmarks.
- Apply it: If you run a multi-model pipeline with a hard-coded router, try scoring each candidate's self-estimated competence per step and routing to the highest, then compare against your static rules.
CausalDS: Benchmarking Causal Reasoning in Data-Science Agents (Leban, Sun / University of Michigan). arXiv
- TL;DR: A benchmark that hands agents a sampled causal model wrapped in a plain-language story, then tests the full ladder from prediction to counterfactuals, with most tasks requiring the agent to write and run analysis code.
- Stat: Tasks span Pearl's full hierarchy, from association through intervention to counterfactual reasoning, with imperfect observations built in.
- Apply it: If your agents do data analysis, add a few counterfactual and confounding questions to your eval set rather than only prediction accuracy. Correlation-level correctness hides causal errors.
🏢 Enterprise in the Wild
- OpenAI is running its own Presence platform on its English-language phone support line, reporting 75% of inbound issues resolved without human help within weeks of launch. It is an early, vendor-reported figure on a narrow channel, but a concrete production data point. VentureBeat
🛠️ Tooling & Ecosystem
- Gemini 3.6 Flash is already available in GitHub Copilot, alongside Google's own Antigravity, AI Studio, and Android Studio. GitHub Changelog
⚖️ Policy & Regulation
- The EU is moving to amend the AI Act as part of its Omnibus simplification package. The changes extend compliance deadlines for high-risk AI systems, expand the lighter-touch small-business framework to firms with up to 750 employees and €150M in revenue, and add rules on AI-generated intimate content. Formal adoption is expected before August 2, the date high-risk requirements would otherwise take effect. Latham & Watkins
📌 Watch List
- Cheaper "flash"-tier models competing on tokens-per-task rather than raw capability.
- Auction and market mechanisms for allocating work across multi-agent systems.
- Causal reasoning benchmarks moving past prediction accuracy.
- Stateless infrastructure for agent tool protocols.