Nous Research just rolled out Hermes Agent v0.7.0, overhauling its memory architecture into an extensible plugin system. It features built-in memory right out of the box, six third-party providers, and the freedom to bring your own backend via hermes memory setup. Giving agents modular, flexible memory just got much simpler.
Anthropic just expanded its computer use feature to Windows across Claude Cowork and Claude Code Desktop. The capability allows Claude to directly navigate desktop apps, browse the web, and fill out spreadsheets. Previously limited to a macOS research preview, Windows users can now let the AI agent handle hands-on desk work.
Pika just launched a beta video chat skill for any AI agent, powered by its new real-time model PikaStream1.0. The tool adds a responsive face and voice while keeping memory and personality intact across calls. Even better: paired with Pika AI Self, agents can execute live tasks without hanging up.
Google just launched Gemma 4, its most intelligent open model lineup yet, released under a commercially permissive Apache 2.0 license. Spanning four sizes from on-device E2B to a 31B dense flagship, the models handle vision, speech, up to 256K context, and agentic workflows, with its 26B MoE outranking models 20 times its size. It basically delivers frontier-level reasoning to local hardware with minimal overhead.
Alibaba just rolled out Qwen3.6-Plus, a native multimodal model built for real-world agentic workflows. It packs upgraded coding execution, sharper visual reasoning, and a 1M context window enabled by default via API, with open-source versions on the way.
Anthropic just published architectural guidance detailing three agent harness patterns that balance intelligence, latency, and cost. The company advises leaning on general tools Claude already masters like bash, stripping away rigid harness orchestration to let Claude manage its own context, and placing static instructions first to maximize prompt cache hits.
Perplexity just rolled out Computer for Taxes inside its Perplexity Computer platform. Built on the Agent Skills protocol, it loads modular tax knowledge grounded in IRS rules to draft federal returns on official forms and audit professionally prepared filings. Internal tests even caught a 67% deduction understatement by an attorney under 2025 overtime rules.
Sakana AI just opened closed beta testing for Marlin, its first commercial AI agent built to act as a virtual Chief Strategy Officer. Given a single prompt, it can autonomously research complex business questions for nearly eight hours, outputting structured presentation slides alongside a dozens-page strategic report. Why it matters: It turns extended inference into full-fledged strategic consulting work.
Moonlake just unveiled a world-building agent that operates directly inside Blender. Instead of one-shot asset generation in chat boxes, it clicks, edits meshes, and adjusts lighting across long-horizon tasks while learning from expert workflows. Why it matters: It shifts 3D AI from isolated generation to executing end-to-end production pipelines in real software.
Perplexity just launched its AI agent orchestrator Computer as a Slack integration, available now in the Slack Marketplace. Teams can tag @Computer in channels or send direct messages to research the web, analyze data, and build apps without leaving their workflow. Tasks and context stay synced between Slack threads and the Perplexity web platform.
Arcee AI just released Trinity-Large-Thinking under Apache 2.0, adding dedicated reasoning steps before responses to sharpen multi-turn tool calling and long-horizon agent stability. The model ranks second on PinchBench right behind Opus-4.6, but its API costs only $0.90 per million output tokens—roughly 96% cheaper.
Elgato just added Model Context Protocol (MCP) support to Stream Deck 7.4, letting AI agents trigger your hardware macros. By running the Elgato MCP Server bridge, tools like Claude Desktop or NVIDIA's Project G-Assist can execute multi-step routines—like launching OBS, dimming lights, and playing audio—purely via text or voice commands. Why it matters: It turns Stream Deck from a manual button board into an execution layer for local AI agents.
Chinese researchers just dropped ASI-Evolve, a 100% open-source agent framework designed to conduct AI research autonomously. It reads papers, designs experiments, runs code, analyzes results, and iteratively improves itself through a closed-loop cycle. Both code and paper are out.
Salesforce just supercharged Slackbot with over 30 new features, turning it into a full enterprise agent hub. It can now transcribe meetings and update CRM records automatically, follow you across desktop apps, and orchestrate other agents via MCP client support. Business+ and Enterprise+ users can toggle it on now.
Perplexity just launched the Secure Intelligence Institute (SII), its new research hub dedicated to AI safety, privacy, and trust. Led by Purdue computer science professor Dr. Ninghui Li, the initiative builds on existing work like BrowseSafe to tackle authentication and defenses for autonomous agents. Why it matters: As AI systems shift from answering questions to taking real-world actions, locking down agentic security is becoming a critical battleground.
H Company just launched Holo3, a suite of computer-use models designed to automate complex desktop workflows. The flagship Holo3-122B-A10B tops the OSWorld benchmark with a 78.85% score using only 10B active parameters, while the smaller 35B version is open-sourced under an Apache 2.0 license.
OpenAI just dropped an official Codex plugin for Anthropic's Claude Code, letting developers invoke a second AI agent without leaving their existing terminal flow. It supports standard code review, adversarial review, and task handoffs by tapping directly into your local Codex CLI and MCP configuration. Why it matters: Dual-agent cross-checking just got seamless—and it even works with free ChatGPT accounts.
Researchers just open-sourced Gen-Searcher, a search-augmented agent framework that scours the web for reference images and facts before generating visuals. By grounding diffusion models in real-time retrieval, it tackles visual hallucinations on obscure entities across both open and closed models.
Anthropic just added computer-use capabilities directly to Claude Code. The CLI agent can now open apps, click through user interfaces, and automatically test the software it just built before handing it back to you. The feature is rolling out as a research preview for Pro and Max subscribers.
Microsoft just upgraded Copilot Researcher with two multi-model tools: Critique and Council. Critique pairs different frontier models from OpenAI and Anthropic to split drafting and fact-checking, boosting benchmark scores by 13.8%, while Council lets you compare divergent answers side by side.
