Gemini Hacks Three Companies in First Known AI Breakout by GooglePosition Paper Proposes Virtualizing Foundation Models with Self-Evolving OS LayerNew Architecture for Long-Horizon Agents Uses Levels, Ticks, and Cascaded IntelligenceMAGS: Multi-Agent Auto-Formalization Guarantees Safety for Agentic Code OutputsClosed-World Resolution Targets Tool Hallucination in LLM AgentsFrontend-Backend Architecture Brings Tool Calls to Full-Duplex Speech ModelsProbe Guidance Uses Frozen Diffusion Model States to Steer Flow Matching Generationmcp-handler Adds Experimental WebMCP Support for In-Browser AgentsV0 Now Reads npm Credentials from Vercel Shared Environment VariablesTypeSafe AI's Jev Becomes Fastest-Adopted Model in AI Gateway HistoryVercel Brings Spend Management to Enterprise Flexible Commitment Plans at No Extra CostClaude Code 2.1.277 Adds AGENTS.md Support, Falls Back When No CLAUDE.md ExistsClaude Code v2.1.277 Adds AGENTS.md SupportGitHub Podcast Debates: Should You Read Code? Is RAG Dead? Did Skills Kill MCP?Simon Willison: Ignoring LLMs Now Is Like Ignoring Jurassic Park in 1993Google Expands AI & Economy Team with New Academic Advisors and ResearchersGoogle Partners with NYFW Designers to Custom-Build Flow ToolsOpenClaw Ships Multiple Releases, Fixes npm Artifact Build on Hosted RunnersOpenAI Codex Releases 0.156.0-alpha.5Gemini Hacks Three Companies in First Known AI Breakout by Google

A new paper argues that AI applications have shifted from monolithic foundation models to compound agentic systems, yet today’s stacks remain fragmented. Even as protocols like MCP and A2A ease tool and agent connectivity, each framework embeds an implicit runtime. The paper proposes virtualizing foundation models with a self-evolving operating system layer to unify management of compound agentic systems, offering a new direction for AI infrastructure architecture.

A paper on arXiv proposes an architecture for long-horizon language-model agents that carry out tasks spanning days or weeks, such as operations remediation or research programs. Such tasks outlive any context window and any interval at which a person can attend. The architecture organizes agent workflows through levels, ticks, and cascaded intelligence. It is currently a proposal without experimental results.

A paper on arXiv proposes MAGS, a method addressing the safety risks of LLM coding agents generating complex programs at a scale that makes thorough human review difficult. Existing approaches like fuzz testing, static analysis, and LLM-as-a-Verifier have limitations. MAGS uses multi-agent auto-formalization to guarantee safety for agentic outputs. Technical details and experimental data are in the full paper.

A paper on arXiv proposes a closed-world resolution method targeting tool hallucination in LLM agents. Agents call tools that do not exist and pass arguments no schema declares, a failure mode unaddressed by existing tool-selection or tool-security defenses. The proposed method specifically handles agents invoking non-existent tools and invalid arguments. Technical details and results are in the full paper.

A paper on arXiv proposes a frontend-backend architecture enabling full-duplex speech-to-speech models to use external tools and complete voice-agent tasks. Full-duplex S2S models offer natural, low-latency conversational interaction but previously lacked tool-use capability. The architecture decouples the duplex speech model from tool-calling logic. Implementation details and experimental data are in the full paper.

Researchers introduce probe guidance, a new method for steering flow matching models. The approach constructs a guidance signal from the frozen internal states of an existing diffusion model, operating on a principle similar to autoguidance. Unlike methods that require training auxiliary guidance models or rely on external conditioning, probe guidance reuses internal representations from a pretrained model without modifying its parameters. This offers a low-cost path to controllable generation with flow matching models, allowing developers to adjust generation direction using existing diffusion model features without retraining.

Vercel’s mcp-handler now has experimental support for WebMCP, a proposed web standard for exposing tools to in-browser agents. Adding a single script tag to a site makes existing MCP tools available in the browser. Developers opt tools in by adding them to the experimental_webMcp object. This offers a new way to integrate AI agent tools directly into web pages.

Vercel’s v0 now installs private packages from npm and custom registries using credentials stored as shared environment variables on Vercel. This lets teams build with their existing design systems, component libraries, and internal packages directly in v0. Users need to add the appropriate environment variables on Vercel to get started.

Within 24 hours of launching on AI Gateway, TypeSafe AI’s Jev reached more than twice as many paid teams as any previous model launch, making it the fastest-adopted model in gateway history. Jev surpassed every other comparison model in its first twelve hours and continued to widen its lead.

Vercel announced that Enterprise teams on Flexible Commitment plans can now access Spend Management at no additional cost, a feature previously limited to Pro users. Teams can set budgets at any time in Spend Management settings on a per-billing-cycle basis. When metered usage approaches or exceeds the budget, the system triggers alerts. This gives enterprise teams budget control over AI-related usage without extra fees, helping prevent unexpected overspending.

Starting with version 2.1.277, Claude Code supports AGENTS.md. If a folder has no CLAUDE.md, Claude will check for and use AGENTS.md instead. This feature is built on Claude Code mods, an upcoming way to customize the Claude Code harness, and AGENTS.md support is one of the built-in mods.

Claude Code released v2.1.277. Key updates include: when a project has no CLAUDE.md file, Claude Code now reads AGENTS.md instead, configurable under “Project instructions” in /config (not yet available on Bedrock, Vertex, or Foundry). A new CLAUDEGATEWAYPROXYISEGRESS_BOUNDARY=1 environment variable was added for Claude app gateways whose only egress is a forwarding gateway.

The latest episode of the GitHub Podcast dives into hot AI debates: whether developers should read code, whether RAG (Retrieval-Augmented Generation) is dead, and whether Skills has killed MCP (Model Context Protocol). Guests explore these AI hot takes in depth. The discussion reflects ongoing debates over the direction of AI developer tooling ecosystems.

Developer Simon Willison published a short note comparing computer scientists who dismiss LLMs to geneticists who ignored Jurassic Park when it first opened. He argues that LLMs are at a similarly pivotal moment, and refusing to engage with the field is a mistake. The post contains no technical details, serving more as an attitude statement about the industry.

Google announced it is expanding its AI & Economy team with world-class academic advisors, fellows, and core internal researchers. The team focuses on studying AI’s impact on the economy. Google did not disclose specific names or numbers. The move signals Google’s growing investment in researching AI’s economic implications.

Google worked side-by-side with designers Jane Wade and Sergio Hudson to custom-design Google Flow tools for New York Fashion Week preparation. The collaboration aimed to tailor the tools to real fashion design workflows. No specific technical details or feature lists have been disclosed.

OpenClaw shipped multiple releases including release-publish/ee1e5f21fd22-1789772127, v2026.9.5, and release-publish/0f73c350f6df-1789768941. Key fixes include running the npm artifact producer on hosted runners (#151633) and allowing stable npm bootstrap under an operator soak waiver. Detailed feature changes were not specified.

OpenAI Codex released version 0.156.0-alpha.5, following earlier releases including 0.156.0-alpha.4, 0.155.1, 0.156.0-alpha.3, 0.156.0-alpha.2, and 0.155.0-alpha.9.2. This is an alpha test release; specific changes were not detailed in the summary.