agent

OpenClaw

2026-10-12 08:17:14

OpenClaw rolls out v2026.10.1 with Skills search and MCP Apps upgrades, but user fatigue is showing

OpenClaw has released v2026.10.1, a sizable update that the project says combines work from 402 contributors across 2,425 pull requests and 3,055 commits. The release introduces two headline changes: agents can now search installed Skills on demand instead of relying on a growing prompt-side directory, and compatible MCP Apps can open directly beside the chat interface for more interactive workflows. The update also expands full backup support to external drives, mounted storage, and Cloudflare R2, while adding scheduled backups, old-backup cleanup, a simpler Mac installation flow, and a series of changes spanning Code Mode, Workboard, automation, model management, and the web interface. At the same time, user reaction shows a different side of the story. Reddit posts and GitHub issue reports point to upgrade fatigue, compatibility headaches, and failed version transitions. Some users say the pace of updates has repeatedly broken customized workflows, even as OpenClaw continues to add features for local-agent use cases.

690

Microsoft

2026-10-12 06:50:48

Nadella says companies should treat AI as an internal risk and keep an emergency stop for agents

Microsoft CEO Satya Nadella said in a lengthy post that companies should handle AI models as part of internal security risk management rather than relying on vendor assurances alone. He argued that models can make mistakes even without malicious intent and can also be exploited by attackers, especially when they are allowed to read sensitive data or carry out critical actions on a user’s behalf. Nadella said the model should be separated from the harness, or agent execution framework, that schedules and runs tasks. Access to data and tools, he wrote, should be controlled by external systems. He also called for tamper-resistant logs for every critical action so human reviewers can inspect what happened. Nadella said chain-of-thought output should be visible, while acknowledging that such reasoning traces may not be reliable. His strongest point was that companies should assume from deployment that a model could be compromised and isolate it accordingly. Authorized personnel, he said, must be able to pause or shut down an agent while it is running, effectively giving it an emergency brake. Microsoft also announced this month that its MXC isolation tool is now generally available on Windows, using external policies to limit which files and network resources an agent can access.

520

Moonshot AI

2026-10-10 04:15:18

Nomura report says Moonshot AI is developing a 5T to 6T-parameter model

ChainCatcher, citing a Nomura report, said AI company Moonshot AI is developing a large model with 5 trillion to 6 trillion parameters. The report said the new system is close to double the size of Kimi K3, which was released in July this year. It also said the main constraint at this stage is a shortage of high-end training compute, and the exact progress of training remains unclear. In the same news brief, Kimi K3.1 was said to be scheduled for release before mid-October, with work focused on improving an agent’s ability to complete long-duration tasks continuously. The team reportedly plans to refine the model through post-training, computer operation records, and real task data. ChainCatcher noted that the information was a relay of the report rather than an official announcement from Moonshot AI.

600

B.AI

2026-10-08 14:25:43

B.AI launches Anthropic’s Claude Haiku 5.5 across API and Web Chat

B.AI said Anthropic’s latest Haiku model, Claude Haiku 5.5, is now live on its platform, with access opened on both API and Web Chat at the same time. The model is aimed at high-frequency calls and latency-sensitive workloads, where speed, cost and intelligence need to be balanced more tightly. According to the announcement, it is suited to sub-agent tasks with clearly defined boundaries. Claude Haiku 5.5 comes with adaptive thinking enabled by default and offers five reasoning intensity levels. It also supports image understanding and tool use, along with a 1 million-token context window and output of up to 128K. B.AI said the model can be used for classification, routing, information extraction, summarization, real-time customer service, code-assisted retrieval and interface automation. Developers and end users can now call Claude Haiku 5.5 directly through the B.AI platform.

470

Sui

2026-10-07 07:10:00

Walrus Console adds portable file and agent memory support with 5 GB free tier

Sui said in a post on X that data should not be locked inside a single application or service provider. It said Walrus Console now supports portability for files and Agent memory, with privacy protection enabled by default and independent verification built in. The post also said users do not need extra syncing or migration to use those features. Walrus Console is offering 5 GB of free storage space, according to the update. The announcement was shared by Sui and framed around giving users a way to keep data portable rather than tied to one platform.

430

ByteDance

2026-10-09 03:44:04

ByteDance merges TraeWork and TraeCode into TRAE, with TraeWork set to shut down on Oct. 23

ByteDance said it is combining its AI office tool TraeWork and coding product TraeCode into a single upgraded platform called TRAE. The two standalone clients will be folded into one entry point, with the new product offering both an Agent mode and an IDE mode. Agent mode is designed to let AI move tasks forward automatically, while IDE mode keeps full code editing and debugging functions in place. According to the company, the upgraded Agent mode can manage multiple projects at the same time, coordinate several AI agents working in parallel, and connect to SSH-based remote development environments. Users can ask the system to handle work starting from requirement analysis through page design, coding, testing, and delivery, then switch to IDE mode whenever manual code changes are needed. ByteDance also said scheduled tasks, template libraries, and office functions from the older products will be integrated into TRAE, with support across desktop, web, and mobile. Official materials state that chat history, account information, local data, login status, and commercial credits will be migrated automatically. TraeWork users must start the upgrade manually in the client, while TraeCode users will receive an automatic update. The company’s documentation adds that TraeWork will stop service on Oct. 23, and users who do not upgrade by then will no longer be able to use it. Some developers have already complained in the official TRAE community that the new version opens in the Agent workspace by default and does not currently offer an option to change the default startup mode to IDE.

650

OKX

2026-10-06 07:41:32

OKX unveils AI ecosystem at Singapore summit and previews AI Bot for natural-language trading

OKX used its OKX NOW summit in Singapore to present a broader AI product stack built around Agent Trade Kit, OnchainOS, Agentic Wallet, OKX AI and the AI Builder Program. According to the company’s on-site presentation, the setup is designed to move AI agents beyond market analysis and into trade execution, while also extending into on-chain asset management, service transactions and developer monetization. OKX also demonstrated an upcoming AI Bot that lets users create personalized trading strategies through natural-language prompts. The ecosystem links the exchange, on-chain infrastructure, wallet services and an Agent marketplace, with AI able to manage assets and execute trades under user authorization. OKX said the framework also supports subscriptions to professional agents and gives developers a way to turn strategies and expertise into income.

390

Perplexity

2026-10-02 02:19:18

Perplexity launches Decisions API and open-sources a 27B decision model

Perplexity has introduced its Decisions API and open-sourced pplx-decider-v1-27b, a multimodal decision model built to return probability distributions across predefined answers instead of generating conventional text. The product is aimed at use cases such as classification, request routing, and choosing the next action for an agent. Perplexity priced the API at $0.04 per million input tokens, with output offered at no charge. In the company’s published benchmark covering 11 tests and 7,210 samples, pplx-decider-v1-27b posted an overall accuracy of 85.71%, compared with 84.51% for Jev and 74.76% for Qwen3.8-27B. The company also noted that the model was not the top performer across every benchmark. Jev scored higher in 6 of the 11 tests, while Perplexity led on FinancialPhraseBank, RAGTruth, TabFact, and Circa. The benchmark figures were released by Perplexity itself and were not produced by an independent third party. The report added that AWS and Cloudflare have also rolled out similar models recently.

1330