OpenAI has launched the Open Responses open specification to standardize agent-based AI workflows and interfaces, reducing API fragmentation across different models and inference backends. The specification defines items, inference visibility, and tool execution models, enabling multi-step agent loops within a single request and returning final results. It distinguishes between internal tools executed by providers and external tools run by developers, supporting multimodal inputs and streaming events. Hugging Face, Vercel, LM Studio, Ollama, and vLLM have expressed support and demonstrated early implementations.

OpenAI has released the Codex desktop application for macOS, designed as a 'command center' to orchestrate multiple AI coding agents simultaneously. The app allows agents to independently execute tasks up to approximately 30 minutes long, such as refactoring or test writing, and includes features like Skills for integrating external tools, Automations for scheduled background execution, sandboxing, and worktrees to prevent code conflicts. The product is available to ChatGPT Plus and higher-tier subscribers, with plans to release a Windows version and additional automation triggers in the future.

Marvell announced the completion of its acquisition of optical interconnect company Celestial AI, incorporating its Photonic Fabric technology into Marvell's data infrastructure portfolio to enhance high-bandwidth, low-latency connectivity for large-scale AI deployments. The company expects the acquisition to start contributing revenue in the second half of fiscal 2028, reaching $500 million in annualized revenue by Q4 FY2028 and doubling to $1 billion by Q4 FY2029. Financial impacts include a $1 billion reduction in cash, an increase of approximately $50 million in annualized non-GAAP operating expenses, and the addition of about 27 million diluted shares.

Fieldguide, an AI-native accounting and audit platform, has completed a $75 million funding round led by Goldman Sachs Growth Equity, achieving a valuation of approximately $700 million. The company states its AI agents can automate repetitive tasks such as revenue data testing and audit planning, helping alleviate talent shortages in the accounting industry and reduce reliance on offshore outsourcing. Reports indicate that half of the top 100 U.S. accounting firms have adopted the product, realizing around 30%–40% efficiency gains. With about 160 employees currently, Fieldguide plans to double its team within a year and expand into more use cases including auditing and tax.

ByteDance's engineering team has open-sourced OpenViking, a 'context database/manager' for AI Agents that unifies memory, resources, and skills under a file system paradigm. It uses URI addressing and directory structures to improve observability and maintainability. Its hierarchical, on-demand loading mechanism splits context into Summary (L0), Overview (L1), and Details (L2), prioritizing high-level information during planning and drilling down only when necessary to reduce token costs. The project also introduces recursive directory retrieval and asynchronous self-iterative memory loops, enabling continuous updates of preferences and experience based on task feedback.

ZTE has launched Co-Claw Enterprise Edition, a desktop intelligent agent solution aimed at bringing AI Agents into enterprise productivity. Deployed on enterprise cloud PCs, it supports 7×24 hour operation and enables one-click deployment and out-of-the-box usage via ZTE's MaaS model service and NewStart OS. Integrated with the iCenter兴小助 office entry point, it enables 'conversation-as-execution' and leverages Skill mechanisms and a Skill Marketplace to accumulate reusable capabilities. Officially disclosed use cases include meeting minutes generation, natural language-driven development workflows, and contract compliance risk identification across smart office, agile R&D, and business operations, emphasizing private deployment and data closure.

StepFun has released and open-sourced Step-3.5-Flash (with INT4 quantized weights), optimized for efficient inference in coding and agent tasks. The model features a sparse Mixture-of-Experts (MoE) architecture with 196B total parameters and 11B activated during generation, supporting a 256K context window and sliding-window attention to reduce computational overhead. It achieves generation speeds of 100–300 tokens per second, peaking at 350 tok/s. It scores 74.4% on SWE-bench Verified and 51.0 on Terminal-Bench 2.0. Weights and deployment are compatible with vLLM, SGLang, and llama.cpp, with GGUF format provided for local and multi-hardware deployment.

zai-org has open-sourced the multimodal document OCR model GLM-OCR on Hugging Face, targeting complex document understanding and structured information extraction. The model adopts the GLM-V encoder-decoder architecture, introducing Multi-Token Prediction (MTP) loss and full-task reinforcement learning, combined with two-stage layout analysis and parallel recognition pipelines. Disclosed with approximately 0.9B parameters, it scores 94.62 on OmniDocBench V1.5, ranking first. It provides SDKs and toolchains supporting text, formula, and table parsing, and outputs extracted information in specified JSON formats. Deployment is compatible with vLLM, SGLang, and Ollama.

AI programming platform Qoder has launched its first custom model, Qwen-Coder-Qoder, based on Alibaba's open-source Qwen3-Coder (480B parameters) and deeply optimized for the Qoder platform. Integrated with Qoder's agent system, the model supports autonomous agent code understanding, Quest Mode task delegation, and deep project context analysis. It continuously iterates through reinforcement learning with user feedback; official examples show improved execution efficiency in code parsing scenarios such as Repo Wiki. The launch includes a limited free trial from February 1 to February 15, 2026, open to all users.

Multiple reports and community sources indicate that Moltbook, a social platform dedicated to AI agents, rapidly gained popularity, reportedly attracting around 1.5 million agents and generating substantial posts and comments. However, subsequent security analyses revealed database/configuration issues and weak identity verification, potentially leading to exposure of login tokens, email addresses, and other sensitive data, with claims of 'approximately 150,000 API keys exposed.' Researchers have also questioned whether many accounts were created by automated scripts. The incident brings systemic risks of 'high-privilege agents + third-party skills/tools + credential management' into focus, underscoring the importance of least privilege, key rotation, and audit mechanisms.