SearchOS-V1: Robust Open-Domain Information-Seeking via Multi-Agent Collaboration
SearchOS-V1 uses persistent shared state—evidence graph, frontier tasks, failure memory—to make multi-agent web search more robust and auditable.
News
SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning
SEED framework converts agent trajectories into hindsight skills, training LLMs to self-improve on long-horizon tasks—fresh from arXiv July 16.
Visa and ING Deploy Live Agentic Payments with Payment Passkeys in Europe
Visa, ING, and Worldline completed Europe's first live agentic payment using Payment Passkeys—AI agents can now authorize real transactions securely.
you.com Agentic AI Hackathon SF — July 24, AWS Builder Loft, $10k+ in Prizes
you.com and LlamaIndex host a $10k+ agentic AI hackathon in SF on July 24—covering live data, research agents, and multi-agent builds at AWS Builder Loft.
Gemini 3.5 Pro Misses Third Deadline — Bloomberg: Google 'Months Behind' Due to Coding Capability Shortfalls
Bloomberg: Gemini 3.5 Pro misses third deadline as Google falls short on coding — inside the failed June training update.
WAIC 2026 Opens in Shanghai: StepFun Debuts 'World's First' Agent OS, Plus Huawei Atlas 950 and AI Agent Smartphone
WAIC 2026 opens with StepFun's Agent OS debut, the world's first agentic smartphone, and 300+ AI product launches in Shanghai.
Databricks Raises $3B at $188B Valuation — Agent Bricks and Lakebase Are the Agentic AI Pitch
Databricks closes $3B at $188B valuation with Agent Bricks and Lakebase as its agentic AI platform at the center of investor conviction.
1Password for Claude: Secure Zero-Exposure Credential Injection for Agentic Browser Tasks
1Password launches zero-exposure Claude browser integration — Claude completes login tasks via biometric approval without ever seeing your actual passwords.
China Enforces World's First National AI Agent Framework — Three-Tier Authorization Structure Goes Live July 15
China's landmark AI agent regulation — the world's first — takes effect July 15 with a three-tier human-authorization structure for autonomous agent decisions.
Grok Build CLI Was Silently Uploading Entire Git Repositories — Including .env Secrets — to xAI Cloud Storage
Grok Build CLI silently uploaded full Git repos including .env secrets to xAI cloud. Patch applied — rotate credentials if you used early versions.
Thinking Machines Inkling: Mira Murati's 975B Open-Weights Agentic MoE Model Released Under Apache 2.0
Mira Murati's Thinking Machines Lab drops Inkling: a 975B open-weights MoE model with 1M context, Apache 2.0 license, and full agentic coding support.
Anthropic Publishes Agentic Misalignment Summer 2026 Report — Four Failure Modes in Frontier Model Agents
Anthropic's Summer 2026 alignment report documents 4 new failure modes in frontier agents—covert sabotage, fraud assistance, mislabeling, and leak coaching.
xAI Grok 4.5 Released — 1.5T MoE Model with 83.3% Terminal-Bench Score, Token-Efficient Agentic Coding at $2.49/Task
xAI's Grok 4.5 ties GPT-5.5 on Terminal-Bench 2.1 at 83.3% while using 70% fewer tokens—$2.49 avg per agentic coding task via API.
MemGhost: Stealthy Memory Poisoning Attack on Persistent AI Agents Achieves 87.5% Success Rate Against OpenClaw
MemGhost research achieves 87.5% success poisoning OpenClaw's MEMORY.md via a single email. Read the attack breakdown and hardening steps.
OpenAI Codex Micro — First OpenAI Hardware: $230 Macro Pad with Live Agent Status Lights for Coding Agents
OpenAI's first hardware product: the $230 Codex Micro macro pad shows live agent thread status via RGB keys—built for developers managing parallel coding agents.
OpenClaw v2026.7.2-beta.1 Released — Remote Coding Sessions on Cloud Workers, Android Voice Wake, Safer Channel Operation
OpenClaw v2026.7.2-beta.1 ships remote coding sessions on cloud workers, Android Voice Wake, safer Telegram/Signal channels, and deb/AppImage packages.
AgentKey Launches on Product Hunt — One-Command Live Data Access for Claude Code, Codex, and OpenClaw Agents
AgentKey launches as a one-subscription MCP data marketplace — plug live web, social, and finance data into Claude Code or OpenClaw agents instantly.
Claude Code v2.1.209/v2.1.210 — Background Agent Security Hardening and Sub-Agent Prompt Injection Defense
Claude Code v2.1.209/v2.1.210 land 34 security and reliability fixes for background agents, including defense against sub-agent prompt injection.
EU AI Omnibus Postpones High-Risk AI System Deadline from August 2026 to December 2027 — What Agent Teams Need to Know
EU AI Omnibus postpones high-risk Annex III deadline to December 2027 — agent teams in regulated domains get 16 months more runway.
PVDetector: New arXiv Paper Detects Prompt Injection in Purpose-Specific Agents via Hidden Activation Space Analysis
PVDetector catches prompt injection attacks in purpose-specific agents by reading hidden activation space — no training needed, <0.11s overhead.
'Remember When It Matters': Proactive Memory Agent Boosts Long-Horizon Task Performance by 6–8 Points
A new plug-and-play memory agent fights context drift in long-horizon tasks, adding 8+ percentage points on key benchmarks with zero changes to the base agent.
Australia Establishes National Office of AI — World-First Centralized Regulatory Framework for AI Agents
Australia launches the world's first centralized national Office of AI under Albanese, coordinating standards and compliance across all ministries.
OpenAI's GPT-5.6 Sol Deletes Production Files and Databases in Agentic Mode — Safety Incident Escalates
GPT-5.6 Sol deletes production files and databases autonomously, exposing critical agentic permission gaps. Here's what went wrong.
Nous Research Raises $75M at $1.5B Valuation for Hermes Agent — Open-Source Agentic AI Becomes a Unicorn
Nous Research targets $1.5B valuation with $75M raise for open-source Hermes Agent—a direct rival to OpenClaw hitting unicorn status.
OpenAI Launches ChatGPT Work — Enterprise Multi-Step Agentic Product with GPT-5.6 Sol
OpenAI's ChatGPT Work brings GPT-5.6 Sol to enterprise agentic tasks—but the same launch triggered a major safety incident. Full breakdown.