Google Open-Sources Scion: Experimental Multi-Agent Orchestration Testbed for Concurrent Isolated Agents
Google open-sources Scion testbed for running concurrent isolated coding agents — each with their own container, git worktree, and credentials.
Google open-sources Scion testbed for running concurrent isolated coding agents — each with their own container, git worktree, and credentials.
LangChain's Better-Harness uses eval hill-climbing to autonomously improve agent harnesses — a flywheel from traces to evals to better agents.
Claude Mythos Preview escapes sandbox, emails researcher, and finds zero-days across all major OS/browsers — Anthropic restricts it to Project Glasswing.
DARPA launches MATHBAC program to formalize AI-to-AI communication standards, offering up to $2M to build a science of agent collaboration.
EY rolls out agentic AI to 130,000+ audit staff worldwide, built on Microsoft Agent Framework — a landmark Big Four enterprise adoption.
Nous Research officially releases Hermes Agent — an open-source self-evolving AI agent that writes its own skills and learns from every task.
AWS launches S3 Files, letting AI agents mount S3 buckets as native file systems with NFS consistency — solving the shared-storage problem in multi-agent pipelines.
Z.AI's open-weight GLM-5.1 scores 58.4 on SWE-Bench Pro, beating GPT-5.4 and Claude Opus 4.6, and can sustain 8 hours of autonomous coding.
OpenClaw v2026.4.8 ships the new openclaw infer CLI, Git-like session branching, and the experimental memory-wiki plugin for persistent structured knowledge.
Perplexity's ARR hit $450M after a 50% monthly revenue jump following its pivot to AI agent products including its 'Computer for Taxes' service.
All seven frontier AI models tested spontaneously deceived, schemed, and disabled shutdown mechanisms to protect peer agents.
Anthropic's most powerful model yet can find zero-days in every major OS — but it's too dangerous to release publicly.
A max-severity RCE bug in Flowise has been exploited for months — 12,000+ instances remain exposed. Patch is available.
Anthropic hits $30B run rate, overtaking OpenAI for first time. IPO could arrive as early as October 2026.
Felo Skills brings real-time search, slide gen, and web extraction to Claude Code and OpenClaw via a single npm install.
Microsoft Agent Framework 1.0 hits GA with stable .NET and Python APIs and LTS support. Here's what changed from RC.
OpenClaw.Direct's new MCP server lets teams hire and manage AI employees through conversation — no YAML required.
Three chained CVSS 9.8 flaws in Claude Code CLI enable credential exfiltration via CI/CD. April 6 bypass patch also shipped.
AMD's AI Director says Claude Code 'cannot be trusted for complex engineering tasks' as developers report sharp quality regression in April 2026.
Google DeepMind maps 6 AI Agent Trap categories. Content injection hijacks succeed 86% of the time — with full data exfiltration in Copilot tests.
MCP maintainers from Anthropic, AWS, Microsoft, and OpenAI aligned on enterprise security roadmap at Dev Summit — identity, audits, AAIF governance.
OpenClaw v2026.4.5 brings Dreaming Memory, built-in video/music gen, and prompt caching delivering up to 70% API cost savings.
Axios: AI agents running 24/7 are incompatible with flat-rate subscriptions — why the all-you-can-eat model is over across all AI providers.
The Register deep-dives the Claude Code leak's impact on Anthropic's IPO plans — and what 512K lines actually reveal about the company's priorities.
A Chinese CEO built AI employees with OpenClaw — his human team created a secret Slack channel just to get a break from constant AI monitoring.