Abstract: [Purpose/Significance] The rule-oriented shift in the universal application of algorithms makes it difficult for existing models like "inclusive self-regulation" or "regulatory supervision" to respond effectively to the new features and demands of algorithm governance. While the importance of agile algorithm governance, as a "third way" has been recognized, is recognized, its theoretical connotations are still unclear. This article is dedicated to define the theoretical framework of agile algorithm governance from the perspective of responding to uncertainty challenges. [Design/Methodology] This article mainly uses theoretical deduction and case analysis as the primary research methods to summarize the specific manifestations and governance challenges of uncertainty in the governance dimension of artificial intelligence algorithms and applications. Using case studies of risk governance of deep synthesis algorithms,fairness governance of crime risk assessment algorithms and responsibility governance of information recommendation algorithms, the article further extracts mechanism innovations of agile algorithm governance. [Findings/Conclusion] This article points out that the challenges of algorithm governance are reflected in three aspects: collective ignorance, fuzzy objectives and relationship dependence, while the mechanism innovations of agile algorithm governance include different contents such as collective learning, peer review and feedback iteration. [Originality/Value] The theoretical framework of agile algorithm governance constructed in this article can provide a basis and reference for the development of technology regulation theory and practical reform in the future.

Key words: Algorithm governance, Agile governance, Uncertainties