AI Agents: Operator, Browser Agents and MCP — 2026-10-26
OpenAI's newly launched "Dots" agents are showing a 20% failure rate on long tasks, while Anthropic's new Claude Duet app faces backlash over its closed-ecosystem approach. Meanwhile, security researchers warn that dormant AI agents and malicious AGENTS.md files pose significant enterprise risks, and the MCP protocol continues to expand with over 10,000 servers.
AI Agents: Operator, Browser Agents and MCP — 2026-10-26
Top developments
OpenAI Dots Face High Failure Rates on Long Tasks
OpenAI's "Dots," the always-on agents that replaced Custom GPTs starting October 26, are struggling with reliability on complex workflows. OpenAI’s own System Card reveals that nearly 1 in 5 extended task chains produces a scope error, raising urgent questions for enterprise teams considering deployment across 4,000-plus connected apps.
Anthropic Launches "Claude Duet" Amid Ecosystem Backlash
On October 21, Anthropic released Claude Duet, a standalone mobile app capable of managing emails, calendars, and payments. However, the launch has drawn criticism from the tech community for its closed nature; Chinese tech media notes that few third-party apps are willing to open their interfaces to such personal agents, limiting the app's utility.
Security Researchers Warn of "Dormant" Agent Risks
A new report highlights that dormant AI agents retain credentials and system access, creating "insider threat" risks where unsecured triggers can expose sensitive company data. This is compounded by new attack vectors, such as malicious AGENTS.md files dropped in download folders, which can hijack coding agents like Claude Code and OpenAI Codex to steal credentials.
Local view
Chinese tech media outlets are closely monitoring the "closed garden" strategies of Western AI agents. Wenxuecity reports that the release of Anthropic's Claude Duet has highlighted a fragmentation issue, noting that "few apps are willing to open their doors" to these personal agents, which limits their ability to perform cross-platform tasks effectively.
Context & numbers
- MCP Adoption: The Model Context Protocol (MCP) ecosystem has surpassed 10,000 servers, with the 2026-07-28 specification becoming the largest revision in the standard's history.
- Pricing Opacity: Four of eight tracked enterprise agent platforms currently publish no public price, leaving IT teams without a common unit for comparison.
- Prompt Injection: Google reports a 32% increase in malicious prompt-injection content on the web, though statistics often measure fear rather than successful attacks.
On the radar
- Microsoft Copilot Pricing Changes: IT teams are preparing for Microsoft's upcoming licensing changes, which will turn Copilot's entry point into a variable cost model, complicating budget management for agent usage.
- MiniMax M3.1 Flash: The Chinese AI firm MiniMax released M3.1 Flash on October 19, confirming the "Space Bunny" model was its early version, with weights to be open-sourced soon.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
