Agent & Model Releases
Hermes Agent Funding. Nous Research is finalizing a $75M round at $1.5B valuation for its open-source Hermes agent, which ships with built-in skills and learns from usage.
Claude Code Browser & Cursor Agent. Anthropic released a Claude Code web browser, while Cursor launched a general-purpose agent, extending coding assistants into broader task automation.
Oak Lab: Self-Learning Agents. Turing Award winner Rich Sutton founded Oak Lab to develop AI agents that learn continuously from experience, aiming for trillion-parameter models that plan and adapt in real time.
GPT-5.6 Sol Spurs Coding Agent Growth. OpenAI's Codex usage surged 10x to 7 million users in six months, with GPT-5.6 possibly overtaking Claude Code in developer mindshare as its efficiency draws users.
Tools & Frameworks
Cicero: Open-Source Voice Agent. Cicero provides bi-directional, self-hosted voice for any agent complying with Agent Client Protocol, working with Hermes, Claude Code, and Codex, and can send Telegram alerts.
DoorDash's Agentic Shopping Assistant. DoorDash detailed its AI assistant's agent architecture using MCP and persistent memory, achieving 24% higher checkout conversion and 17% larger baskets.
AI Agents Verify Crypto Code. Microsoft's SymCrypt project uses AI agents to scale formal verification of cryptographic algorithms, releasing verified Rust code for SHA-3 and ML-KEM.
Coding Agents Boost Open Source. Simon Willison's Datasette project saw a record spike in code activity, aligning with releases of advanced coding models like Opus 4.8 and GPT-5.6.
Research & Innovation
Robotic Agent OS with Lifelong Memory. ABot-AgentOS introduces a runtime layer with multimodal graph memory for robots, enabling long-horizon tasks and self-evolution from failure diagnostics.
Visual Language Navigation Model. ABot-N1 decouples reasoning from control using pixel goals as a universal interface, achieving robust navigation across diverse embodied tasks.
World Models Simulating Reality. Practitioners from MIT, Runway, and World Labs discuss world models that simulate physical environments, starting with creative tools and expanding to broader AI.
Hy3 Model and Efficient Inference. Tencent's 299B MoE Hy3 model is now supported in llama.cpp, and Colibri streaming optimization allows it to run on as little as 10GB of combined RAM/VRAM.
On-Device AI Advances. PrismML compressed Qwen 3.6 27B to run on an iPhone 17 Pro, while Apple plans an M7 Ultra chip with up to 1.5 TB unified memory for local model inference.
Industry & Business
Apple Sues OpenAI for Trade Secret Theft. Apple's lawsuit alleges OpenAI poached employees and stole hardware secrets, including coaching interviewees to bring product samples and bypass security.
Nadella Warns of Double Payment for AI. Microsoft CEO Satya Nadella cautions that companies pay AI providers with both money and proprietary data, calling out the irony of banning distillation while training on public data.
Anthropic Talent Influx. Tom Blomfield, Mike Krieger, and Andrej Karpathy have joined Anthropic, drawn by the formative moment at the frontier of LLMs.
Anthropic Extends Fable 5 Access. Amid a pricing war with OpenAI's GPT-5.6 Sol, Anthropic extended free Fable 5 access for subscribers through July 19 and raised usage limits.
Chinese Models Gain Enterprise Traction. Companies are adopting Chinese open-weight models like Qwen to cut costs; meanwhile, a German consortium released Soofi S 30B, topping benchmarks in English and German.
Security & Alignment
Defenders Use Prompt Injection. Tracebit researchers showed that placing forbidden prompts alongside sensitive data can cause attacking AI agents to shut down, turning the attackers' tool against them.
That's everything for this day – about a 5-minute read.