Nexus: стартап Singularity Escape привлёк посевной раунд на ИИ-платформу для командных агентов

Singularity Escape closed a seed round of tens of millions of yuan on August 2, 2026, to develop Nexus — an AI operating system for collaborative agent teamwork. The round was led by Stellar Capital and Shuimu Ventures, with accelerator Miracles Startup participating as a co-investor.

Why Single Agents Fail in Team Settings

A typical scenario in a modern team: the product manager hands requirements to one agent, the developer to a second, the marketer to a third. A few hours later, each agent has a finished result and full context for its own task. But when the three parts need to be combined, it turns out that no one knows what the neighboring agent actually did, what decisions it made, or why.

• Round closed in August 2026; amount — tens of millions of yuan
• Lead investors: Stellar Capital and Shuimu Ventures; co-investor — Miracles Startup
• Founder Xue Chuanyi — PhD from Tsinghua University, publications at NeurIPS and CJA
• Product: Nexus — an OS for collaborative work with AI agents
• Key concept: a unified organizational state graph instead of isolated chats

Xue Chuanyi, CEO of Singularity Escape, calls this the "collaboration gap": the human once again becomes an information intermediary — copying context between agents, explaining the same things in different windows. The agent remains a "smart loner" rather than a full-fledged team member.

"Today an agent is a capable individual, but it has not yet entered the organization," —

Xue Chuanyi, CEO of Singularity Escape, in an interview with 36Kr.

How Nexus Works

Nexus represents people, agents, tasks, knowledge, and tools as interconnected nodes in a unified graph. Each task contains not just a text instruction, but a complete package: goal, responsible party, dependencies, access rights, current status, proof of completion, and acceptance criteria.

When a new agent joins a project, it sees the context of previous tasks — like a new employee who has received a full onboarding. All results are returned to the shared graph and become the starting point for the next executor. This moves the agent from the category of a personal tool to that of an organizational participant who can be delegated to, granted rights, and replaced without losing context.

Nexus uses a two-level graph architecture: within the agent system, nodes include the goal, planning, specialized skills, memory, and tools; at the organizational level — people, agents, tasks, and knowledge. This structure makes it possible to precisely track where an improvement occurred and which parts of the system it affects.

Self-Learning Under Evidence-Based Control

Self-learning in Nexus works in three stages: real feedback from tasks → independent testing of candidate improvements → adoption of changes through a governance mechanism.

The system can propose adjustments — changing how context is organized, model routing, or workflow configuration. But no change goes directly to production: first comes independent testing, comparison by effectiveness, stability, cost, and risks.

"Self-evolution is not unlimited autonomy.

It is stricter boundaries of evidence and governance," — states official Singularity Escape materials. The system is required to answer: why a change is needed, what exactly changed, whether it is better, and whether it can be rolled back.

According to Xue Chuanyi, the long-term value of Nexus accumulates not in model parameters, but in organizational state — the history of tasks, decisions made, and the evidence base. This asset will not be reset when the underlying language model is updated.

What This Means

Singularity Escape is betting that business value is created not by a single powerful agent, but by a team system that accumulates organizational intelligence. If the Nexus concept is realized in production conditions, the startup could occupy a niche as an infrastructure layer for companies already building multi-agent workflows.

Frequently Asked Questions

Who founded Singularity Escape?

Xue Chuanyi is a scientist of the "post-2000" generation (born after 2000), with a bachelor's degree and PhD from Tsinghua University in reinforcement learning and multi-agent systems. He is the first author of publications at the NeurIPS conference and in the CJA journal; he interned at SenseTime and Jizhi Robotics, where he developed algorithms for autonomous vehicles.

How large was the seed round?

According to 36Kr, the round amount is "tens of millions of yuan" (千万级), with the exact figure not disclosed. At current exchange rates, this corresponds to approximately $1.4–10 million depending on the precise amount.