Microsoft Research just introduced MMCTAgent, an open-source framework built on AutoGen to reason across long videos and massive image libraries. Instead of relying on one-shot answers, it deploys Planner and Critic agents to break down queries, trigger computer vision tools, and cross-check evidence. On the MM-Vet benchmark, it pushed GPT-4V's accuracy from 60.20% to 74.24%.
Landing AI just set a new state of the art on DocVQA with 99.16% accuracy, beating both Qwen2-VL and human baselines. The team decoupled parsing from querying: its DPT-2 model extracts layout and text into structured data once, leaving downstream questions entirely to an LLM. The parse-once approach cuts costs and makes every answer fully traceable.
Dharmesh just launched Company Research agent, an AI tool designed to automate company and startup deep dives. It pulls firmographics, executive rosters, web traffic, and G2 reviews into a single report that can be customized, shared without a login, and synced straight into HubSpot.
ElevenLabs just launched Scribe v2 Realtime, a low-latency speech-to-text model built specifically for live voice agents and meeting assistants. It transcribes audio in under 150 ms across 90 languages and uses next-word prediction to cut perceived lag. This brings conversational AI much closer to natural, instant banter.
ByteDance’s Volcano Engine just debuted Doubao-Seed-Code, a coding agent priced at an introductory 9.9 yuan ($1.30) for the first month and 40 yuan thereafter. It handles up to 256,000 words per query, boasts a 78.8% score on SWE-bench Verified, and is already integrated into the Trae IDE to replace Anthropic's Claude.
Terminal-Bench just announced two major updates: Harbor, a tool built for running sandboxed agent rollouts at scale, and Terminal-Bench 2.0. The new benchmark cranks up task difficulty and verification to give developers a much more rigorous test of agent capabilities.
Crypto insiders just broke down the surge of AI trading bots, warning investors that automated tools are far from foolproof money-makers. Experts from Glassnode and Cindicator pointed out that AI excels at crunching data rather than magical market forecasting, and general chatbots fall flat in quant trading. Today's bots function more like 24/7 interns that still require human oversight.
OpenAI just detailed its multi-layered strategy to defend AI agents against prompt injection attacks. As autonomous models browse the web and handle sensitive data, hidden malicious instructions inside web pages or emails pose serious security risks. OpenAI is deploying instruction hierarchy training, automated monitors, sandboxing, and sensitive-action confirmations. Why it matters: Preventing agents from going off-script is the prerequisite for letting AI handle real-world work.
Perplexity just rolled out a completely reimagined Comet Assistant to all users. Built into its Comet browser, the agentic assistant can now handle multi-tab workflows simultaneously, making tasks like cross-referencing spreadsheets and travel searches much smoother. Perplexity says internal tests show a 23% performance gain, backed by tighter user permission controls before it touches your browser.
Google Labs just expanded Opal, its no-code AI mini-app builder, from 15 countries to over 160 worldwide. The tool lets anyone build multi-step workflow automations, custom content generators, and quick MVPs without writing code. Basically, turning loose ideas into custom AI productivity tools is now accessible to almost anyone.
Google just upgraded Structured Outputs in the Gemini API, adding full JSON Schema support that works out-of-the-box with Pydantic and Zod. The update brings support for unions, recursive schemas, and strictly preserves key ordering across Gemini 2.5 models. For developers building multi-agent systems, it means finally putting an end to broken JSON responses in production.
Coinbase just launched Payments MCP, a framework that equips models like Claude, Gemini, and Codex with crypto wallets and stablecoin rails without requiring API keys. Running locally on desktop, agents can now pay for compute or paywalled data via the x402 protocol, all within user-set spending limits.
Opera just expanded its Neon AI browser with ODRA, a dedicated deep research agent. It breaks down complex queries, runs parallel searches across multiple models like OpenAI and Google, and compiles the results into detailed summaries. Why it matters: It marks a shift toward multi-agent orchestration directly inside browser workspaces, though Neon is still invite-only.
Microsoft just upgraded Dynamics 365 with autonomous agents designed to tackle core operations across sales, customer service, and supply chains. Highlights include a Sales Close Agent that proactively flags pipeline risks and a Quality Evaluation Agent for real-time interaction audits. Starting in late November, select Premium subscribers will also receive 1,000 pooled Copilot Credits per user each month.
OpenAI has quietly recruited over 100 former investment bankers for a secretive initiative dubbed Project Mercury, paying $150 an hour to train its models on complex financial modeling. The goal is to automate the grueling Excel tasks traditionally handled by junior analysts, from corporate restructurings to IPO valuations.
OpenAI just launched ChatGPT Atlas, a brand-new web browser built around its AI assistant. Atlas features browser memories to recall context across sites you visit, alongside an agent mode that can autonomously click tabs, compare research, and book items directly on web pages. It is rolling out globally on macOS first, with Windows and mobile versions on the roadmap.
Anthropic just launched Claude for Life Sciences to assist researchers from early drug discovery to regulatory compliance. Powered by Claude Sonnet 4.5, which outperforms human baselines on lab protocol QA, the platform adds connectors to PubMed, Benchling, and BioRender alongside modular agent skills. Why it matters: LLMs are moving beyond generic office tasks to run end-to-end scientific workflows.
Anthropic just launched Claude Code on the web, letting developers delegate coding tasks straight from the browser without touching a local terminal. Tasks run in parallel inside isolated cloud sandboxes, automatically submitting PRs across multiple GitHub repos—and there's even an iOS preview for coding on the go. Why it matters: It turns routine bug fixing and repo maintenance into hands-off background work.
Napster just launched View, a 65g holographic display designed to make AI agents feel physically present on your desk without headsets or glasses. It connects to M1 or newer Macs via USB-C to display 3D specialists for coding, writing, and screen sharing. It is available for $99 standalone or free with an annual plan.
Elon Musk just announced that X will replace all its heuristic recommendation algorithms with Grok within 4 to 6 weeks. The AI will process over 100M posts and videos daily to boost quality content from small accounts, and users will be able to tweak their feeds simply by talking to Grok.
