Researchers developed WaLLM, a general-purpose LLM chatbot, and deployed it on WhatsApp to study user behavior. The study found that users primarily used WaLLM for health and well-being advice, and that engagement features had varying adoption rates and associated user patterns.

Researchers propose a method called 'grafting' to apply pre-training interventions to AI models without requiring a full post-training run, reducing iteration time and improving model stability.

Mingbird is a local-first agent harness designed for small open models to complete real tasks. It addresses issues such as tool prefill overflows, self-correction divergence, and task abandonment by introducing ten mechanisms, including a byte-level net-zero prefill budget and signature-level loop detection. Mingbird outperforms other harnesses on various benchmarks, achieving 0.886 overall on LRAB and 0.856 on τ^2-bench.

YouRA is a persistent-state architecture for evidence-traceable autonomous research agents. It integrates a Verification State Architecture, an Independent Controller, and Stateful Reflection to track hypotheses, gates, and evidence pointers, and improve the reliability and transparency of research agents.

Cloudflare launches Cloudflare OHTTP Gateway, a paid add-on to enable Oblivious HTTP (OHTTP) traffic for app backends, providing privacy-preserving infrastructure for developers.

Cloudflare has launched eight major updates to its observability platform, including logs, traces, analytics, and alerts, with simpler pricing and a unified API. This will help developers and operators gain a more complete view of their applications and resolve issues more efficiently.

Cloudflare introduces Web Search API via AI Gateway, enabling agents to search the internet and access live information, improving the accuracy and relevance of AI models.

Rogo's Vercel setup automates AI agent deployments in 5 minutes, enabling 73,000+ monthly deployments and zero manual triage during production incidents.

Vercel releases a Python AI SDK for working with Jev, a new type of AI model that can be used as a classifier, allowing users to make narrow decisions without needing to train a model for a specific domain.

A new open-source memory and judgment layer for AI coding agents records development as an append-only log of typed events and provides a deterministic summary through the Model Context Protocol. This allows for project-specific rationale retention and repeat failure prevention.

Researchers introduce Galahad, a memory layer for LLMs that enables stateful inference by storing and reusing model key-value state, reducing computation costs and energy consumption.

Cloudflare's AI Search is now generally available, providing a fully managed index and retrieval pipeline with multimodal support, including native image embeddings and OCR for PDFs. This allows for more accurate and detailed searches, especially for products, screenshots, and scanned documents.

Cloudflare challenges developers to build the next-gen Git platform for agent-based development, leveraging their Artifacts versioned filesystem and Workers.

Cloudflare OS is a managed platform for AI agents that provides a custom workspace for organizations, allowing them to connect to their data and systems. It has been expanded with new features, including mounting Git repos, working with code, and exporting work in various formats.

Cloudflare introduces Workers KV Instant, a new mode for Workers KV that provides instant availability and low-latency access to data, powered by the Quicksilver key-value store. It offers 100 times faster p99 reads and immediate updates, with no need to wait for a TTL to expire, making it ideal for reading data in the hot path of applications.

Cloudflare announces two new open-source language models, EuroLLM and Apertus, developed by European research institutions, to promote choice and sovereignty in AI. The models are available on Workers AI and designed to work with any model, focusing on accessibility and multilingual support.

Grist is an open-source coding harness that uses the Opencode v2 framework, leveraging the jev model for task routing and integrating with cheaper models like deepseek and Opus 5.5. It also incorporates the Sol-Pi methodology for cost-cutting and loads engineering standards through Doctrine injection. The tool has a real control plane for escalation, permission, and verification hooks, and can be run with a key from Openrouter or Vercel AI gateway. It also supports connecting to Meta Muse for coding tasks.

Allen Institute for AI announced the release of Olmo-core 3, an open and scalable training infrastructure for large-scale models. This update brings improved performance, increased flexibility, and better support for distributed training. Olmo-core 3 is designed to accelerate the development and deployment of large models, making it easier to build and train AI agents.

ThinQuant is a new method for efficient rotation learning in large language models (LLMs). It reduces the computational cost of rotation learning by introducing a data selection procedure and an exact reduction of the optimization problem. This allows ThinQuant to scale to large architectures and achieve comparable performance to state-of-the-art methods like DartQuant and GPTAQ+QuaRoT.

This paper proposes Environment Steering, a technique to improve agent utility and safety by enforcing safety as the agent runs and steering it toward safe alternatives when violations occur. It uses a declarative policy and context-specific feedback to track data flows and steer the agent toward safe trajectories.