1. NVIDIA Releases Nemotron 3 Embed Open Model Collection
NVIDIA has released the Nemotron 3 Embed collection, featuring three open transformer encoder checkpoints: an 8B BF16 model, a 1B BF16 model, and a 1B NVFP4 model optimized for Blackwell hardware. All three models support bidirectional attention masking and a maximum sequence length of 32,768 tokens. The 1B models were developed using NVIDIA ModelOpt Neural Architecture Search for pruning and distillation from the 8B teacher. Additionally, NVIDIA released an optimized Rust-based NIM microservice for the 1B model that performs on par with or better than the vLLM checkpoint on GB200 and RTX PRO 6000 hardware.
- • The Nemotron-3-Embed-8B-BF16 model ranks first on the Retrieval Embedding Benchmark (RTEB) with an average NDCG@10 score of 78.46.
- • The collection includes an 8B BF16 model, a 1B BF16 model, and a 1B NVFP4 model optimized for Blackwell hardware.
- • The 1B NVFP4 model provides up to 2x higher throughput on Blackwell hardware compared to the BF16 version while retaining over 99% of its retrieval accuracy.
- • All models are released under the OpenMDW License Agreement, version 1.1, and support a maximum sequence length of 32,768 tokens.
Developers can deploy highly accurate, open-weight embedding models with up to 32k context for production RAG, agent memory, and code retrieval.
2. Trellis.cpp Fixes Bugs in GGML-Ported TRELLIS.2 3D Generation
The developer of trellis.cpp has announced that critical bugs in the GGML-ported TRELLIS.2 image-to-3D generation pipeline have been fixed. The asset quality of the ported model is now on par with the reference model, enabling high-quality, open-source 3D generation on local hardware via GPU or CPU. The raw engine is available on GitHub and can be integrated with Lemonade for an enhanced experience and optional text-to-3D cascade.
- • Bugs in the GGML-ported TRELLIS.2 image-to-3D generation pipeline have been fixed, bringing asset quality on par with the reference model.
- • The tool enables open-source 3D generation locally on hardware with sufficient GPU power or via CPU.
- • The raw engine is hosted on GitHub and can be used with Lemonade for an integrated experience and optional text-to-3D cascade.
Developers can now run high-quality, open-source image-to-3D generation locally on consumer hardware using an optimized GGML port.
3. Capital One Open-Sources VulnHunter AI Security Tool
Capital One has released VulnHunter, an open-source agentic AI security tool available on GitHub under an Apache 2.0 license. VulnHunter scans source code for vulnerabilities, maps potential exploit paths, and proposes targeted code fixes before production. The tool utilizes an attacker-first forward analysis method, tracing potential exploit paths starting from system entry points like APIs or file uploads. It also includes a built-in falsification engine designed to disprove potential vulnerabilities before presenting them to developers, reducing false positives.
- • Capital One released VulnHunter, an open-source agentic AI security tool, under an Apache 2.0 license.
- • The tool scans source code, maps exploit paths using an attacker-first forward analysis method, and proposes code fixes.
- • A built-in falsification engine attempts to disprove potential vulnerabilities to minimize false positives.
- • VulnHunter currently operates on Anthropic's Claude Opus 4.8 model within a Claude Code environment.
Developers can integrate an open-source, agentic security scanner into their CI/CD pipelines to automatically trace exploit paths and generate targeted code fixes.
4. Brex Open-Sources CrabTrap Network Proxy for AI Agents
Brex has released CrabTrap, an open-source network governance tool for AI agents. Operating as an HTTP/HTTPS proxy, CrabTrap intercepts network traffic to enforce security policies using a combination of deterministic static rules and an LLM-as-a-judge for unfamiliar requests (which typically account for less than 3% of traffic). The tool is framework, language, and API-agnostic, requiring only standard proxy environment variables. To mitigate prompt injection, CrabTrap structures requests as JSON objects to escape user-controlled content before evaluation.
- • CrabTrap is an open-source HTTP/HTTPS proxy that intercepts network traffic to enforce security policies on AI agents.
- • The tool uses deterministic static rules and triggers an LLM-as-a-judge for less than 3% of total traffic.
- • It is framework, language, and API-agnostic, requiring only HTTP_PROXY and HTTPS_PROXY environment variable configuration.
- • CrabTrap structures requests as JSON to escape user-controlled content and mitigate prompt injection risks.
Developers can secure their agent deployments by intercepting and filtering outbound network requests through a framework-agnostic proxy.
5. GitHub Releases Copilot SDK for Custom App Integration
GitHub has made the Copilot SDK available, enabling developers to integrate GitHub Copilot agents into their own custom applications and developer tools. This SDK provides the necessary interfaces to leverage Copilot's agentic capabilities within external workflows, expanding where and how developers can deploy Copilot-powered assistance.
- • The Copilot SDK is available in a public GitHub repository.
- • The SDK enables the integration of GitHub Copilot agents into custom applications and developer tools.
Developers can embed GitHub Copilot's agentic capabilities directly into their proprietary developer tools and custom internal applications.
6. Anthropic Reverses Claude Code Auto-Continue Feature
Anthropic released Claude Code version 2.1.198 on July 1, 2026, introducing an undocumented auto-continue feature for the AskUserQuestion tool that proceeded after 60 seconds of inactivity. The feature included an internal "away from keyboard" (AFK) designation, a countdown timer, and analytics tracking. Following user feedback, Anthropic reversed this behavior in version 2.1.200 on July 3, 2026. Because the Claude Code source code is not public, users had to diff compiled binaries to identify these changes.
- • Anthropic introduced an undocumented auto-continue feature in Claude Code version 2.1.198 that proceeded after 60 seconds of inactivity.
- • The feature was reversed in version 2.1.200 on July 3, 2026, following user feedback.
- • Users can manage the auto-continue behavior or disable the auto-updater using environment variables like CLAUDE_AFK_TIMEOUT_MS and DISABLE_AUTOUPDATER.
Developers using Claude Code should update to version 2.1.200 or configure environment variables to prevent unexpected automated tool execution.
7. Anthropic Details Large-Scale Code Migrations with Claude Code
Anthropic has detailed its internal methodology for running large-scale code migrations using Claude Code. The process relies on a six-step strategy that includes creating a rulebook, analyzing dependencies, and stress-testing translation rules. To ensure migration efficiency and code quality, Anthropic deploys multiple agents to iteratively translate, review, and fix code, emphasizing the use of adversarial reviewers and mechanical verification.
- • Anthropic uses a six-step process for large-scale code migrations using Claude Code.
- • The strategy involves deploying multiple agents to iteratively translate, review, and fix code.
- • The process includes creating a rulebook, analyzing dependencies, and stress-testing translation rules.
- • Anthropic utilizes adversarial reviewers and mechanical verification to improve migration efficiency.
Developers can adopt Anthropic's proven multi-agent workflow patterns to automate complex, large-scale codebase migrations with higher reliability.
8. Intuit Rebuilds Agent Architecture Twice in Four Months
At VB Transform 2026, Intuit's VP of AI, Nhung Ho, revealed that the company rebuilt its agent architecture twice within four months. Intuit initially moved from specialist agents to a central orchestration layer, which failed due to compounding errors from natural language handoffs. The company then successfully transitioned to a skills and tools-based architecture in 60 days. This modular approach allows individual components to be reused across products and shifts engineering focus to system evaluation. Intuit is also testing a feature allowing human support staff to join active AI agent conversations with full context.
- • Intuit abandoned a central orchestration layer after three months due to compounding errors caused by natural language handoffs between agents.
- • The company rebuilt its system into a skills and tools-based architecture, completing the second rebuild in 60 days.
- • The new architecture allows individual components to be reused across products, shifting engineering focus to system evaluation.
- • Intuit is testing a feature allowing human support staff to join active AI agent conversations with full context.
Developers can avoid costly architectural dead-ends by bypassing natural language agent handoffs in favor of a reusable, skills-and-tools-based design.
9. Tech Leaders Identify Legacy Infrastructure as AI Agent Bottleneck
At VB Transform 2026, infrastructure leaders from LinkedIn, Walmart, and Zendesk discussed the challenges of scaling AI agents. LinkedIn addressed slow Kubernetes provisioning by moving to pre-provisioned container pools and reduced hallucinations by shifting to 80% deterministic code, reserving LLMs only for reasoning. Zendesk prioritized data pipeline investments over feeding large conversation histories directly into context windows, while Walmart implemented governance to manage agent duplication. Both LinkedIn and Walmart developed internal AI gateways to maintain model and vendor independence.
- • Infrastructure leaders identified legacy infrastructure rather than AI models as the primary bottleneck for scaling agents.
- • LinkedIn reduced agent hallucinations by shifting to 80% deterministic code and using LLMs only for reasoning.
- • LinkedIn addressed slow Kubernetes provisioning by moving from on-demand container creation to pre-provisioned pools.
- • LinkedIn and Walmart built internal AI gateways to maintain model and vendor independence.
Developers can design more reliable agent architectures by adopting pre-provisioned container pools, deterministic code paths, and internal AI gateways.
10. Harness Handbook Maps Agent Behavior to Code Implementation
The Harness Handbook has been released as a behavior-level map designed specifically for coding-agent harnesses. It bridges the gap between plain-language questions regarding execution, permissions, and safety, and the concrete prompts, tools, state logic, configuration, and telemetry required to implement them. This mapping helps developers systematically design and audit the operational boundaries of their AI agents.
- • The Harness Handbook is a behavior-level map designed for coding-agent harnesses.
- • It connects plain-language questions about execution, permissions, and safety to specific prompts, tools, state logic, configuration, and telemetry.
Developers can use this handbook to translate abstract agent safety and execution requirements into concrete, production-ready code and prompts.
11. MongoDB Releases Tutorial for Agentic Event Venue Operator
MongoDB has released a reference tutorial for building an agentic event-venue operator demo. The architecture utilizes MongoDB Atlas as a unified layer for operational records, semantic memory, visual documents, and agent state. The application stack integrates FastAPI, Claude Vision for Vision RAG, Voyage AI embeddings, and LangGraph for agent orchestration, with optional Langfuse integration for observability. The demo features a four-tab UI and hybrid retrieval combining vector similarity with lexical scoring.
- • MongoDB released a tutorial for building an agentic event-venue operator using MongoDB Atlas, Voyage AI embeddings, and LangGraph.
- • The architecture uses MongoDB Atlas as a unified layer for operational records, semantic memory, visual documents, and agent state.
- • The application stack includes FastAPI, Claude Vision for Vision RAG, and optional Langfuse for observability.
- • Prerequisites include Python 3.12 or later, a MongoDB Atlas cluster with Vector Search, and API keys for Anthropic and Voyage AI.
Developers can follow this reference architecture to build multimodal, stateful agents that leverage MongoDB Atlas as a unified database and memory layer.
12. Observer Screen-Watching App Integrates Local LLMs
Observer, an open-source application that uses LLMs to monitor screen activity and send notifications, has transitioned to a SignPath Foundation FOSS project under the AGPLv3 license. The application now supports local inference through browser-based transformers.js, Ollama, bundled llama.cpp, vLLM, or any OpenAI-compatible endpoint, ensuring screen data remains on the user's local machine. Additionally, setup has been simplified to a single natural language command.
- • Observer is an open-source, AGPLv3-licensed screen-monitoring application that supports local LLM inference.
- • The application supports local inference via transformers.js, Ollama, bundled llama.cpp, vLLM, or OpenAI-compatible endpoints.
- • Setup has been simplified from a manual multi-step process to a single natural language command.
- • Screen data remains entirely on the user's local machine to preserve privacy.
Developers can build privacy-first screen-monitoring and automation workflows using fully local models and a simplified setup process.