Snowflake Explores Agent Context Management
Snowflake AI Research describes a context-efficiency approach spanning Snowflake CoCo and Snowflake CoWork, combining dynamic tool loading, parallel execution of independent tool calls, and compression of large tool outputs. The company also presents governed semantic views, including automated generation and review through Semantic Studio, as a way to supply business context. In Snowflake's internal SQL-repair benchmark, CoCo with Sonnet 5 reportedly achieved an 86% pass rate versus 72% for Claude Code with Opus 5, while reported cost fell 33% to $0.298 per trial.
The report frames agent efficiency as a problem of context curation rather than context-window size, combining runtime orchestration with reusable business definitions for more focused and cost-aware agents.

2026 年,智能体将在企业级应用中取得哪些实质性突破?点击下载"《2026 年 AI 与数据发展预测》白皮书,获悉专家一手前瞻,抢先拥抱新的工作方式!
你的 Agent 并不缺信息,它只是被信息淹没了。
每一项工具定义、每一次查询结果、每一轮对话,都会增加 Agent 工作时需要处理的上下文。在一项持续时间较长的调查任务中,这些额外开销会不断累积。即使早期步骤中的信息已经不再相关,Agent 仍可能在后续每一轮交互中反复携带这些内容。
这就是上下文膨胀(Context Bloat)。它不仅会增加 token 消耗,还可能让 Agent 更难聚焦于真正重要的信息。
单纯扩大上下文窗口,并不能从根本上解决这一问题。更有效的方法是管理哪些信息进入上下文、哪些信息需要保留,以及 Agent 在开始执行任务之前需要掌握哪些知识。
#Snowflake#Snowflake CoCo#Snowflake CoWork#Semantic View Autopilot#Semantic Studio#Claude Code#Sonnet 5#Opus 5