Let Agents Self-Evolve in Collaboration, Tsinghua Post-00s PhD Receives Ten Million Yuan in Financing | 36Kr Exclusive
奇点逃逸完成千万级种子轮融资，由星连资本与水木创投领投、奇绩创坛跟投，创始人薛传奕为清华00后博士，研究方向为强化学习与多智能体 公司正在研发AI原生团队协作操作系统Nexus，核心理念是让人、Agent、任务、知识和工具基于同一份组织状态持续协作，解决当前Agent协作中的"协作断层"问题 Nexus采用图结构作为底层组织方式，将目标、规划、执行、记忆、工具等建模为可追踪的节点，使改进可以定位、验证和回滚 自进化机制分为三个环节：真实反馈驱动、独立评测筛选、治理机制采用，强调严格证据和治理边界而非无限自主权 长期定位是"AI时代的组织基础设施"，关注组织整体处理复杂目标的能力而非单个Agent功能
Analysis
TL;DR
- 奇点逃逸完成千万级种子轮融资，由星连资本与水木创投领投、奇绩创坛跟投，创始人薛传奕为清华00后博士，研究方向为强化学习与多智能体
- 公司正在研发AI原生团队协作操作系统Nexus，核心理念是让人、Agent、任务、知识和工具基于同一份组织状态持续协作，解决当前Agent协作中的"协作断层"问题
- Nexus采用图结构作为底层组织方式，将目标、规划、执行、记忆、工具等建模为可追踪的节点，使改进可以定位、验证和回滚
- 自进化机制分为三个环节：真实反馈驱动、独立评测筛选、治理机制采用，强调严格证据和治理边界而非无限自主权
- 长期定位是"AI时代的组织基础设施"，关注组织整体处理复杂目标的能力而非单个Agent功能
Why It Matters
This startup addresses a critical gap in the current AI Agent ecosystem: while individual Agent capabilities are rapidly improving, there is no robust infrastructure for multi-Agent collaboration within organizations. The concept of "collaboration fault lines" (协作断层) accurately describes the pain point where human workers become information relays between isolated Agent sessions. For AI practitioners, this highlights that the next competitive frontier is not model capability but organizational integration and continuous learning from real production feedback.
Technical Details
- Nexus Architecture: Built on a dual-layer graph structure — at the Agent-internal level, nodes represent goals, planning, capabilities, execution, memory, tools, and verification; at the organizational level, nodes represent people, Agents, tasks, and knowledge. This enables traceability of improvements and their impact across the system.
- Collaboration Model: Shifts from session-based information organization to organization-state-based collaboration. Work items are represented as structured objects containing goals, owners, dependencies, permissions, current status, execution evidence, and acceptance results — not just natural language prompts.
- Self-Evolution Loop: Four-stage cycle: (1) real collaboration produces tasks and feedback, (2) independent evaluation filters valid improvements, (3) verified capabilities enter the next collaboration round, (4) stronger Agents expand organizational collaboration boundaries. Candidate improvements (context organization, model routing, task decomposition, workflow configuration) must pass evaluation before production deployment.
- Governance Framework: Self-evolution requires strict evidence and governance boundaries — the system must answer why changes are made, what was changed, whether it genuinely improves performance, whether existing capabilities are harmed, and whether rollbacks are possible.
- Data Asset Philosophy: Accumulates a different type of data asset than model parameters — recording not just what answers were produced but why work was completed in a particular way, derived from real task judgment and feedback.
Industry Insight
- The "collaboration fault line" problem signals that the Agent market is approaching an inflection point: individual Agent performance gains will face diminishing returns without organizational infrastructure. Investors and builders should prioritize platforms that enable multi-Agent coordination and persistent organizational memory.
- The self-evolution framework with independent evaluation and governance gates represents a more mature approach than "free self-modification" narratives. Organizations adopting Agent systems will demand auditability and rollback capability — startups that build these safeguards into their architecture will have a competitive moat.
- The dual graph structure design suggests that the next generation of AI-native collaboration tools will blur the boundary between collaboration software and ML training pipelines, creating a new category of "learning organizations" where execution feedback directly drives capability improvement.
Disclaimer: The above content is generated by AI and is for reference only.