XAI’s recently released Grok Bot has become very popular for persistent, cloud-hosted multi-agent workflows. Instead of one bot, you create different connected bots to accomplish different tasks in multi-agent way; you can think of them as different AI employees serving different roles for you. Since good AI design patterns get copied, others are replicating how Grok Bot works.Nous Research released Bot Mode for the Hermes agent. The mode allows Hermes agents with specialized roles to operate persistently and coordinate automated tasks:Your agent profiles become a series of named Bots. Each Bot has its own role, model, memory, skills and profile picture; Bots can use any model and even communicate with each other.This adds a great feature to the already great Hermes Agent.CopilotKit released OpenBot as an open implementation of the persistent multi-agent bot pattern. It follows the pattern of cloud-hosted, always-running agents used by Grok Bot and similar systems and It developers create specialized agents that remain available and coordinate work across tasks.Ornith released the open-source Ornith-1.5 model family, consisting of a 9B parameter dense model and 35B and 397B parameter MoE (mixture-of-experts) models. These models are based on Qwen 3.5 and Gemma 4 base models and were post-trained through a self-improvement process that generated tasks, constructed agent scaffolds and self-evaluated work. Ornith-1.5 397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, benchmark results comparable to DeepSeek-V4-Flash-0731. Ornith-1.5-35B-A3B is an efficient AI model for local AI coding agents, on par with Muse Glimmer 30B.Xiaohongshu’s dots studio released dots3-note preview, an open-weights MoE model with 280B total and 16B active parameters and a 512,000 token context window. It targets long-horizon agentic tasks, and it is multimodal, accepting text, images, video and audio input, with text-only output.Ant Group’s InclusionAI released the Ling-3.0 MoE model family: Ling-3.0-tiny has 7.9B total parameters with 1.3B active parameters per token; Ling-3.0-flash has 124B total and 5.1B active parameters, while achieving mid-tier results, comparable to MiniMax-2.7. InclusionAI released pretrained, intermediate-training and merged checkpoints, with quantized versions on HuggingFace intended for local deployment.Cartesia released the Sonic-3.6 text-to-speech model, the streaming speech model that supports 44 languages and can produce audio in under 90 milliseconds latency. Sonic-3.6 reached first place on Artificial Analysis’ Provider Voice and Controlled Voice leaderboards, providing an independent indication of its speech quality.Alibaba launched HappyShrimp 1.0 for end-to-end music generation, which converts prompts describing a genre, emotion or story into complete songs, generating the lyrics, melody, arrangement, vocals and final audio production. Users can also provide their own lyrics. Independent comparisons of its musical quality and consistency against systems such as Suno or Udio are not yet available.Liquid AI released new four-bit LFM2.5 checkpoints for edge device models. The 230-million to 2.6B parameter LFM2.5 models were trained with quantization-aware distillation to recover approximately 97% of their BF16 benchmark performance with low memory footprint and higher throughput.xAI released Grok Build to all users. Grok Build generates working websites, applications, games and dashboards inside Grok conversations on the web, iOS and Android. Users can revise projects conversationally, publish them to dedicated URLs and connect generated applications to xAI APIs for Grok chat, image and voice capabilities.TrueFoundry open-sourced TrueForge, an open agent harness that supports models from over 20 other providers, along with MCP servers, sandboxed execution, human approvals, persistent sessions, context management and self-hosted deployment. TrueFoundry reported 30% to 75% lower task-completion costs than Claude Managed Agents in its tests, but those results depended on model selection, workload and the company’s own evaluation methodology.Figure 2. The TrueForge Open Source Agent Harness architecture.Superwhisper released S1-mini as an open-weight transcript-cleaning model. The 0.6B parameter Qwen3 fine-tune is specialized for English transcript cleanup; it converts raw speech-recognition output into punctuated, formatted written text and can run locally, avoiding sending data to a server. S1-mini has been released in standard and GGUF formats for local use.Amazon Bedrock expanded access to Grok 4.6 and OpenAI’s GPT-5.6 models. xAI made Grok 4.6 generally available through Bedrock, while AWS added cross-region inference for GPT-5.6 Sol, Terra and Luna across more than 25 regions.Adobe made its Firefly AI audio tools generally available. Firefly’s AI-based Generate Music, Generate Speech and Generate Sound Effects tools produce background music, voiceovers and precisely timed sound effects inside Adobe’s creative platform. Adobe said the commercially released tools were trained on licensed and public-domain material. The new free Firefly AI Assistant provides limited daily generations.Figure 3. Adobe Firefly Audio features are now available, allowing users to add jingles and songs in AI video generations.Salesforce launched Slack Code for Slack-based collaborative agentic programming. Dedicated code channels allowed teams to assign work to agents such as Claude Code and Devin, examine plans and code changes, run previews, provide feedback and approve deployment from a shared Slack workspace. This moves AI coding activity out of private terminals and into auditable team channels, a feature that other agents and applications such as Claude Tag provide.Anthropic added a /design command to Claude Code. The research-preview command creates and revises Claude Design artboards from Claude Code’s command-line interface or the Claude desktop application. It integrates visual interface design into a coding workflow, allowing developers to iterate layouts before implementing them in code.Cursor launched Origin, a Git-compatible code-hosting platform that competes with GitHub. Origin integrates repositories, code review, continuous integration and preview deployments with Cursor’s coding agents and was designed to accommodate the high commit volumes produced by multiple agents working in parallel. Cursor also expanded its cloud agents to be able to modify and test repositories in isolated environments and continue working while users were offline.Chroma introduced Foundation as a persistent-memory service for AI agents. Foundation stores information and outcomes from previous sessions as a self-improving wiki and make them retrievable across tools and AI agents. This avoids having to force each agent session to rediscover the same context and procedures and solves the episodic agent memory challenge.OpenAI added Apple Messages integration and expanded Computer History in ChatGPT Work. The Apple Messages plugin allows ChatGPT and Codex on Apple Macs to search for iMessage, SMS and RCS conversations and prepare or send messages, with recipient and message approval enabled by default.OpenAI previewed Private Safety Processing for zero-data-retention customers. In this system, customer content remains on customer-controlled infrastructure or in storage encrypted with customer-controlled keys, and OpenAI receives only narrowly defined safety signals when automated systems detect a risk. The system analyzes patterns across related API interactions to detect misuse or agent behavior that could not be recognized from a single request, while keeping the underlying prompts and responses inaccessible to OpenAI.Modular released the complete Mojo compiler and toolchain as open source. Mojo is designed as a general-purpose systems language for CPUs, GPUs, AI accelerators and high-performance AI software, combining Python-like syntax with lower-level control. The Apache 2.0 release with LLVM exceptions opened the compiler, tooling and language implementation after Qualcomm completed its acquisition of Modular.Ramp launched Router, a unified API for accessing AI models. The service allows businesses to switch among multiple AI models from different providers while monitoring token spending, latency and fallback attempts. Router is limited to U.S. customers for now.Meta expanded its Pocket vibe-coding application to users across the United States. Pocket uses Meta’s Muse Spark model to create small interactive games and other experiences from natural-language prompts and distribute them through a social feed where others could play, share and remix them.Apple Music is preparing for mandatory labels for AI-generated and materially AI-created tracks. The planned “Made With AI” labels will replace Apple’s earlier voluntary approach and require music distributors to identify qualifying releases.OpenAI paused frontier AI model Astra development after new cybersecurity warnings arose in preliminary testing of its unreleased Astra model, which suggests it possesses critical cyber capabilities. OpenAI paused reinforcement-learning work on its models for two weeks to focus on strengthening its safeguards. They are adding stronger sandbox and network isolation, trajectory-level monitoring, and continuous security testing.MIT researchers documented “attribution decay” in diffusion models. Researchers trained 24 diffusion-model ensembles on datasets ranging from 256 to more than 160,000 images and found that removing an individual image, person, or artist increasingly failed to change the generated output as datasets grew. The finding applied to diffusion models and individual-level causal attribution; it did not establish that training datasets were legally irrelevant, eliminate occasional memorization, or demonstrate the same effect in LLMs.Google DeepMind published an overview tracing 15 years of game-based AI research in games from Atari reinforcement learning to the development of SIMA 2. The foundational techniques developed through game research, including RL, have contributed to development of reasoning AI models and specialist AI models such as Alpha-Fold as well as broader scientific breakthroughs.An example of AI advancing science is how Claude designed experimentally validated protein binders. Claude autonomously designed 1,320 protein candidates across 15 targets by orchestrating specialist protein-design and structure-prediction models. Laboratory testing found 354 binders against 14 of the 15 targets with interpretable results, representing an overall hit rate of about 27%.The downstream benefits of AI can be seen in the news that Moderna and Merck reported positive Phase 3 results for an individualized mRNA cancer therapy. AI and automated algorithms reportedly contribute to the sequence design and were part of the design pipeline for this mRNA therapy that involves individualized medicine, showing that AI can accelerate developing new cures.Stripe agreed to acquire OpenRouter, a gateway that routes AI requests across more than 400 models from over 80 providers, giving developers a unified interface for selecting models, managing availability and controlling token expenditures. Reports valued the transaction at over $8 billion. Stripe said it planned to combine model routing with its token-billing and financial infrastructure while retaining OpenRouter’s multi-model approach. OpenRouter CEO says their mission continues.Cursor announced that its acquisition by SpaceX had closed and that its team would join SpaceXAI.Data-labeling startup Micro1 grew its gross annual run rate from $100 million to $500 million over an eight-month period driven by surging demand for AI training data , showing the expanding market for specialized training data as top AI labs scale their models. The company relies on a network of domain experts for model evaluation and reinforcement learning and is expanding into automated synthetic data generation and robotics pre-training datasets.Ramp released transaction data showing that OpenAI began narrowing Anthropic’s market lead among U.S. business users during the third quarter. Anthropic held a lead over OpenAI in corporate adoption earlier in the year, but OpenAI accelerated its growth following the release of its GPT-5.6 Sol model.OpenAI cofounder and president Greg Brockman significantly expanded his operational purview to become second-in-command behind CEO Sam Altman. Brockman assumed leadership over the company’s product strategy, super app efforts, and scaling division as OpenAI reorganized ahead of an anticipated initial public offering.OpenAI joined the PORTS-Pike Technology Data Center project at the former Portsmouth uranium-enrichment site in Ohio. The planned campus would provide as much as eight gigawatts of AI computing capacity under a 20-year lease.Filmmakers Matti Haapoja and Sam Kolder faced public backlash after posting promotional videos for AI video platform Higgsfield and Seedance 2.5. Fans and fellow creators criticized the lack of disclosure of potential paid partnerships and the normalization of generative AI tools in professional filmmaking. The controversy adds to ongoing debates regarding AI-generated content, and it also highlights how AI’s public-acceptance problem has become a business PR problem.OpenAI announced AI Futures, a new initiative led by Dean Ball focused on long-term policy, economic integration, and institutional governance for advanced AI. In launching OpenAI’s Strategic Futures team blog AI Futures, Dean Ball argues that autonomous systems could allow governments and powerful institutions to operate military, administrative, and economic machinery with less dependence on human labor and public cooperation.He proposes balancing power among institutions while preserving individual access and autonomy rather than relying on either centralized control or unrestricted decentralization.the structural challenge posed by advanced machine intelligence to free society is likely not the most radical decentralization of power imaginable. Instead, it is seeking to establish and preserve the right balance of power, such that no single actor, or small set of actors, can dominate the rest.Striking this balance will require clear-eyed thinking about the reality of this technology. A world in which any individual, however malicious, can trivially cause great harm to thousands or millions of other people is not a world which has struck the right balance of power. Yet neither have we struck the right balance of power if individuals lack broad access and control over almost all aspects of the tools they use to express their freedom. No one company or oligopoly should determine or control the economy or the basic architecture of human society. – Dean Ball