Thousand-Agent Orchestration Becomes the Selling Point as Security and Cost Emerge as Adoption Barriers

Highlights

• Anthropic Fable 5.2 coordinates over a thousand subagents, builds a prototype operating system in eight days
• Meituan open-sources LongCat-2.5: 1.6 trillion parameters, million-token context, built for long-horizon agents
• Baidu Dazi wins three bids in a row; office Agent MAU up 1,567.5% month-over-month in September

Model layer: "multi-agent collaboration" as the capability benchmark

Anthropic on Nov 9 launched Claude Fable 5.2 and Claude Mythos 5.2. Both are built on the same underlying model: the former targets general users and ships with built-in cybersecurity and biosecurity safeguards, while the latter is available only to vetted cybersecurity and life-science organizations. On pricing, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, 20% below the $10/$50 of 5.1; generation speed is up to 30% faster, and fast mode can push that to 2.5x at a higher price. (Source: ITHome)

More notable is the task shape seen in pre-release testing: the model coordinated over 1,000 subagents to build from scratch, in eight days, a prototype operating system that boots on a laptop; and it used about 1,000 agents to complete a Lean formalization of Fermat's Last Theorem in 30 hours, against a previous record of 11 days. This shows vendors are now using "orchestrating large numbers of subagents to complete long-horizon tasks" as a way to demonstrate model capability, rather than single-turn Q&A. (Source: ITHome)

On the domestic side, Meituan on Nov 9 announced the open-sourcing of LongCat-2.5 under the MIT license, supporting commercial inference deployment and further training. The model has 1.6 trillion total parameters, activates about 48 billion parameters per token, and supports a one-million-token context. It is positioned as a native multimodal model for long-horizon agent tasks, with architecture and deployment co-optimized for domestic compute chips constrained by memory and bandwidth. (Source: ITHome)

Product layer: the battle for the entry point shifts from conversation to "getting things done"

Baidu Dazi ranked first in three government and enterprise projects between Nov 5 and 6: CITIC Securities, State Power Investment Corporation Digital Technology, and General Technology Digital Technology successively published candidate lists, with five, four, and three vendors shortlisted respectively. Baidu Dazi is positioned as the unified entry point for the Agent era, able to orchestrate capabilities including Baidu Search, Miaoda, Shengsuan, and Famou. QuestMobile data for September 2026 shows its multi-platform MAU grew 1,567.5% quarter-over-quarter, the fastest growth among office Agent products. (Source: Leiphone)

Meta, meanwhile, is putting its personal agent Muse on the wrist: the iOS version of Muse now supports Apple Watch, letting users type, dictate, or make voice calls, with the yeti mascot as the core visual element of the watch face. Muse already has a standalone app and is integrated into Instagram, Facebook, and other Meta apps. (Source: ITHome)

Adoption barriers: verification, auditing, and cost

As capabilities expand, constraints are being pushed onto the table. An IEEE Spectrum article discusses using zero-trust architecture to manage agent activity, with the core principle "never trust, always verify": when an orchestrator dispatches multiple agents to execute tasks, how do you prove instructions were followed, and where is the evidence? The article also mentions context-recovery attacks—information that was asked to be deleted may be elicited again by later prompts. (Source: IEEE Spectrum)

On V2EX, a developer reported that since Oct 20, nearly all of their private repositories have seen malicious commits, including credential-stealing scripts disguised as woff2 files, shell executed via postinstall, and branches created for no apparent reason. The user suspects these operations were carried out en masse by AI agents, citing extremely fast commit speeds and scripts clearly generated from package.json context; after ruling out PATs, SSH keys, 2FA logins, and VPS channels, they pointed suspicion at an Alibaba Cloud Yunxiao GitHub OAuth Access Token. The allegation is currently a personal investigation conclusion, and no platform response has been seen. (Source: V2EX)

Cost is also a real constraint. In another discussion in the same community, a developer handed two projects of about 10,000 lines of code each to codex, and found that switching to another agent required rescanning the repository every time—slow and token-hungry. "My wallet can't take it anymore," the developer wrote, seeking a workflow that balances quality, speed, and cost. (Source: V2EX)

When model vendors use "thousand-level subagents" to prove the upper limit, enterprises and individual developers face two other questions: who keeps verifiable logs of agent behavior, and who pays the bill for rescanning the repository every time.