01AgentsProductsingle source
Beatsprom has released the Qwen-3.8-Coder-Stateless-MCP-Agent, a production-grade autonomous tool-calling agent for Claude 5, GPT-6, and DeepSeek workflows, with a reported 26.4% increase in BFCL tool-calling accuracy and 28.6% increase in SQL injection defense. The agent is available for use with PEFT and can be run locally with Ollama. Commercial tooling and a full 25 MCP stack are also available for purchase on Gumroad.
Provenance — who else covered this
02ResearchInternalssingle source
CAIS and Scale released HLE-Diamond, a cleaned subset of Humanity's Last Exam, while OpenAI released MentalHealthBench, built with input from 80+ clinicians. Other releases include OpenRSI-Index v0.1, WorkspaceBench, and updates to various benchmarks and environments, such as Google's RRSI, Salesforce's RIVER, and NVIDIA's Skill2Env. Research findings include Microsoft's discovery that k agents sharing a directory can match 4k independent agents on ARC-AGI-3 and Stanford's demonstration of a self-organizing team hitting 66.7% on a task.
Provenance — who else covered this
03AgentsInternalssingle source
Recent research has made significant progress in agentic reinforcement learning, enabling large language models to reason and act autonomously over long time horizons. New frameworks and techniques, such as ToRL, AgentGym-RL, Agent-R1, and AgentRL, have been proposed to train LLM agents with RL, improving their performance on various tasks and environments. These advancements have the potential to unlock more capable and generalizable AI systems.
Provenance — who else covered this
04AgentsProductsingle source
DAPLab has released several benchmarks for evaluating AI agents, including LAKEQA, Live Kernel Crash Resolution, and BranchBench, with more in development, such as SRE-Bench and ConcurBugBench. These benchmarks test various aspects of AI agents, including their ability to discover and reason over structured datasets, diagnose and resolve real-world failures, and support agentic workloads. The benchmarks are designed to provide verifiable outcomes and contamination-resistant evaluation, and are updated regularly.
Provenance — who else covered this
05ModelsInternalssingle source
DeepSeek AI introduces the DeepSeek-V4 series, a new generation of language models that achieve significant improvements in efficiency and performance, particularly in long-context scenarios. The series includes two models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, which incorporate innovative architectural upgrades and optimization techniques. The models have been pre-trained on large datasets and demonstrate state-of-the-art performance on various benchmarks.
Provenance — who else covered this
06AgentsProductsingle source
A comprehensive guide to five agent orchestration patterns - orchestrator-worker, swarm, mesh, hierarchical, and pipeline - has been released, covering their architecture, trade-offs, and real-world implementations. The guide helps practitioners choose the right pattern for their multi-agent systems based on task structure, latency requirements, scale, and observability needs.
Provenance — who else covered this
07ModelsProductsingle source
H Company has released Holotron-12B, a multimodal computer-use model that achieves high throughput inference and strong agent performance, and is now available on Hugging Face. The model is post-trained from the NVIDIA Nemotron-Nano-2 VL model and demonstrates significant improvements over the base model on various benchmarks.
Provenance — who else covered this
08ResearchProductsingle source
The credibility of LLM benchmarks is compromised due to contamination and saturation, making it difficult to rely on leaderboard scores. A guide is provided to help readers critically evaluate benchmarks, considering factors such as governance type, contamination, and saturation. The goal is to move towards dynamic, contamination-resistant evaluation methods that prioritize real-world performance over raw scores.
Provenance — who else covered this
09ModelsInternalssingle source
BFL released an open-weights 7B world-action model, FLUX 3 Action, which takes the top spot on RoboLab, beating the previous best open model by 6.1 points with 56% fewer parameters. Other notable releases include CLM-8B, tev1-4B, and Cua-S1-4B-0.2 models, as well as Apple's LensVLM and inclusionAI's Ming-Image-0.1-Design. Additionally, Inferact open-sourced a TPU megakernel for Kimi K3, and several infrastructure updates were announced, including Prime Sandboxes, ClusterMAX 3.0, and a 25T-token pipeline.
Provenance — who else covered this
10AgentsProductsingle source
OpenAI has released the Agents API in public beta, allowing developers to rent the harness behind Codex and ChatGPT for Work, with a managed orchestration layer priced toward zero, and including features like automatic compaction and parallel subagents. The API has some limitations, including US-only data residency and token-and-tool-metered pricing.
Provenance — who else covered this
11AgentsInternals5 sources agree
Asymptote Labs has open-sourced Beacon, a self-improving memory layer that uses Jev to evaluate and classify agent runs at scale, allowing for the creation of a shared knowledge layer across multiple agent harnesses. Beacon preserves full session history and helps identify valuable lessons to promote, review, or discard.
Provenance — who else covered this
12AgentsInternals4 sources agree
Researchers introduce Harness-Zero, a method that distills an optimized agent harness into the model, resulting in higher macro task success rates. Additionally, several other advancements are reported, including the release of SmolDataEnvs, a dataset of verifiable data-science RL environments, and the development of an AI-run lab and the SciUniverse benchmark by C5R.
Provenance — who else covered this
13CodingProduct2 sources agree
DHH, a prominent skeptic of AI-written code, has publicly changed his stance, citing significant productivity gains from using AI-generated code, and now advises developers to stop writing code by hand, while others discuss the implications of this shift on workflow and collaboration with AI models. The numbers behind DHH's reversal are striking, with a 5x increase in production code output in a single month, and a 99% reduction in CPU usage and 95% reduction in memory usage after rewriting HEY's mail server in Rust.
Provenance — who else covered this
14ResearchProduct2 sources agree
John Platt discusses Google's Empirical Research Assistance (ERA), a tool that uses AI to solve scientific problems, and its applications in tackling climate change, including reducing the effects of condensation trails from airplanes. ERA is an open-source implementation that can be used with any large language model.
Provenance — who else covered this
15ResearchProductsingle source
A guide to understanding LLM benchmarks, their limitations, and how to evaluate vendor claims, highlighting the importance of context and governance fit in model selection. Atlan's research shows that improving governed metadata context can significantly improve AI accuracy.
Provenance — who else covered this
16AI securityProductsingle source
Codex CLI 0.155.0 introduces experimental voice chats and Touch ID verification for MCP requests, while also hardening Windows/WSL sandboxes and protecting shell snapshots. The update adds a biometric gate for MCP tool calls, giving agents more surface and moving approval closer to the OS.
Provenance — who else covered this
17ModelsInternalssingle source
Contrastive-LM introduces CLM, an open-weights, self-hostable alternative to TypeSafe AI's Jev, offering the same functional primitives with improved latency and fine-tunability, but trailing in zero-shot broad knowledge and context budget. CLM achieves state-of-the-art verifier performance on certain benchmarks when fine-tuned on agent trajectories.
Provenance — who else covered this
18On-deviceInternalssingle source
Liquid AI DSpark achieves up to 3.13× decode speedup with MLX on M5 Max, while other projects like GLM-5.3 and Pruna few-step LoRAs also show significant efficiency gains, and Google launches Project Suncatcher with TPUs in orbit.
Provenance — who else covered this
19ResearchProductsingle source
Researcher claims OpenAI modified SWE-bench to favor their models, and a paper by Singh et al reveals big LLM vendors are gaming Chatbot Arena by submitting multiple variants and fine-tuning models. This compromises the validity of these leaderboards, which are considered more meaningful than others due to their complexity and real-world relevance. The issue highlights the difficulty in maintaining fair evaluations in the AI research field.
Provenance — who else covered this
20CodingProductsingle source
Opus 5.5 is receiving praise from practitioners for its ability to handle long-horizon orchestration without breaking a sweat, with many reporting that usage limits now feel effectively unlimited. Builders are testing cross-CLI patterns and reporting efficiency gains, with some noting that the model is surprisingly good and feels much smarter. However, a migration debate is emerging, with some developers switching back to Claude Code after time on the ChatGPT/Codex app.
Provenance — who else covered this
21ResearchProductsingle source
Radical Numerics is developing Genomic Language Models (GLMs) to tackle biological problems, with applications in biosecurity and beyond. The company's co-founder, Eric Nguyen, discusses the potential of GLMs to unlock biological intelligence and the need for defense to keep pace with advancements in the field. The company's models have already shown promising results in generating DNA and protein sequences, and are being developed to think in the language of DNA, with potential applications in fields such as epigenetics and natural language processing.
Provenance — who else covered this
22AgentsProductsingle source
Anthropic's Claude Skills library offers 372 AI skills across 20 domains, while Azure Agent Skills provides 193 skills across 19 categories, and AgentSmith introduces a model-agnostic operating harness for verification, with builders positioning skills as complementary to MCP, raising questions about standardization and vendor lock-in. Meanwhile, Notion and other companies are exploring skills tooling and agent-to-agent comms protocols.
Provenance — who else covered this
23AgentsProductsingle source
Smolagents is a lightweight, code-first agent framework that enables developers to build powerful agents with minimal abstractions, comparing favorably to alternatives like LangChain and Microsoft Agent Framework. It supports model-agnostic design, modality-agnostic inputs, and sandboxed environments for security. The framework is ideal for single-agent, short-horizon reasoning, tool-heavy workflows, and prototyping, but lacks persistent memory, a limitation addressed by integrating Mem0.
Provenance — who else covered this
24ResearchProduct4 sources agree
Jev offers a new interface for evaluators, allowing for parallel structured judgments, which can be more efficient and cost-effective than traditional LLM judges, and a complete evaluation workflow using Jev and Comet's Opik platform is available for exploration and use.
Provenance — who else covered this
25BusinessBig picture3 sources agree
Adrian Sanborn, CEO of Endura Therapeutics, discusses how AI is transforming front-line scientific research, enabling faster and more agile experiments, and improving access to expertise, with examples from his company's work in biotech and CRISPR research. He highlights the role of 'navigators' who use AI to drive better decisions and faster processes, and the potential for AI to accelerate scientific progress by making expertise more accessible and scalable.
Provenance — who else covered this
26On-deviceProduct2 sources agree
Backlot, a local emulator for enterprise SaaS APIs, and open-supermarkets, an open-source grocery interface, simplify testing and integration for agent builders, while prompt patches and doc-native versions address agent halting issues.
Provenance — who else covered this
27AgentsInternalssingle source
AG2 v1.1.0 adds a TypeSafe provider for decision-only agents and updates MCP surfaces to MCP 2.0, while other developments include a citation bug fix in Gemini 3.8, the introduction of Talkie, a 13B LLM for generalization research, and RedSOC, an adversarial eval framework showing high success rates in indirect prompt injection and corpus poisoning.
Provenance — who else covered this
28BusinessBig picturesingle source
Governor Gavin Newsom signed a package of bills requiring large AI data centers to cover their own electricity infrastructure costs, introducing new permitting processes and water-use disclosure requirements. This change creates a separate utility rate class for data centers and gives communities greater control over project approvals, affecting the build-versus-buy calculus for agent deployments.
Provenance — who else covered this
29ResearchProductsingle source
A new paper from Cohere Labs and other institutions criticizes the Chatbot Arena leaderboard for allowing model providers to anonymously preview and cherry-pick their models, leading to biased scores. The paper also highlights unfair sampling rates and removal rates. The authors propose more transparency in the testing process. Meanwhile, OpenRouterAI LLM rankings are suggested as an alternative evaluation method.
Provenance — who else covered this
30ResearchInternalssingle source
A new study reveals that popular LLM evaluation platforms, such as Chatbot Arena, are vulnerable to changes in top model rankings when a small fraction of preference data is removed, highlighting the need for more robust evaluation methods. The researchers propose a fast and efficient method to assess the robustness of LLM rankings to worst-case data-dropping, which can help identify influential preferences and improve the reliability of leaderboard rankings.
Provenance — who else covered this