Tracked 282 · Curated 15
A confluence of events in late 2025, including the release of Anthropic's Opus 4.5 and OpenAI's GPT 5.2, marked an inflection point for agentic AI in early 2026. This period heralded the dawn of the agentic AI era, with OpenClaw emerging as a key development.
The LightSeek Foundation has released TokenSpeed, an open-source LLM inference engine under the MIT license, engineered for agentic workloads. It aims to match TensorRT-LLM-level performance by optimizing for both high per-GPU TPM and user TPS. TokenSpeed's architecture features a compiler-backed parallelism mechanism, a high-performance scheduler, KV resource reuse restrictions, a pluggable layered kernel system supporting heterogeneous accelerators, and SMG integration. Benchmarks on NVIDIA B200 using SWE-smith traces and the Kimi K2.5 model show TokenSpeed outperforming TensorRT-LLM by approximately 9% for agentic coding workloads above 70 TPS/User.
Mozilla detailed how it used a preview of Anthropic's Claude Mythos to find and fix hundreds of vulnerabilities in Firefox. This shift from AI bug reports being 'unwanted slop' is attributed to improved LLM capabilities and Mozilla's refined techniques for harnessing them. The number of security bugs fixed in Firefox jumped from around 20-30 per month in 2025 to 423 in April, including a 20-year-old XSLT bug and a 15-year-old bug in the <legend> element.
Cloudflare is reducing its global workforce by over 1,100 employees, citing fundamental changes in how the company operates due to the rise of AI. The company states this restructuring is to optimize operations and accelerate innovation for the agentic AI era, not a cost-cutting measure or performance assessment. Departing employees will receive significant severance, including pay and healthcare support through the end of 2026.
Microsoft's Azure Kubernetes Fleet Manager addresses the complexity of managing thousands of Kubernetes clusters at scale. It enables teams to group clusters into stages for controlled, sequential rollouts and updates, reducing manual intervention. The solution leverages Cilium Cluster Mesh for seamless cross-cluster connectivity and unified management, essential for distributed workloads like AI.
Parloa leverages OpenAI models to power scalable, voice-driven AI customer service agents. This enables enterprises to design, simulate, and deploy reliable, real-time interactions.
PhysForge, a two-stage framework for physics-grounded 3D asset generation proposed by Tencent researchers, has been accepted at ICML 2026. It uses a VLM architect for hierarchical blueprints and diffusion with KineVoxel Injection to create simulation-ready assets, trained on 150K PhysDB.
OpenAI has released its official command-line tool, openai-cli, enabling developers to call its APIs directly from the terminal without writing SDK code. The project is open-sourced on GitHub (openai/openai-cli) under the Apache 2.0 license and can be installed via Homebrew or Go. This tool supports functionalities such as calling response APIs, generating structured outputs, image generation/editing, transcription, TTS, and managing projects and API keys.
Claude Code has announced increased usage limits: the 5-hour limit for Claude Code is doubling for Pro, Max, Team, and seat-based Enterprise plans. Peak hour limit reductions are removed for Pro and Max Claude Code usage, and API rate limits for Opus models are substantially raised.
ByteDance Seed has released PV-VAE, a predictive Video VAE model that trains on partial context to reconstruct and forecast future frames. The model improves latent diffusability with 52% faster convergence and achieves 34.42 FVD gains over Wan2.2.
The Max Agency podcast interviewed Alex Shevchenko, Head of Applied Research at Ramp Labs, discussing the construction of Ramp Sheets, their internal Agent Inspect, and more. The interview is available on YouTube, Apple, and Spotify.
GPT image 2.0, released two weeks ago, continues to reveal surprising new capabilities. Users have discovered its ability to generate text posters and its powerful anime-style image generation. It can even produce images based on named IPs without requiring reference images.
Chinese AI lab Deepseek is nearing a funding round that could value it at approximately $45 billion, according to the Financial Times. The round is being led by China's state chip fund.
The Mythos model is confirmed to be non-marketing hype. It's a general-purpose model that's good at finding exploits, and similar models are expected from OpenAI & Google, with open models to follow in 8 months.