从 MCP 到 REPL：下一代 Agent 的标准执行层 | From MCP to REPL: The Standard Execution Layer for Next-Generation Agents

​📍 Tiered Classroom 阶梯教室

​我们认为，下一代 Agent 的执行层既不会停留在 MCP，也不会止步于 CLI，而会走向以 REPL 为代表的新形态。MCP 解决了工具接入的标准化，CLI 释放了系统环境的组合能力，但复杂工程现场需要的是更连续、更可表达、更贴近运行时的执行层：一个让 Agent 能够在真实系统中探索、验证、执行，并沉淀能力的工作界面。

​本次分享将围绕这一判断展开：Agent 的执行层正在从“接口调用”和“命令执行”，走向更高抽象层级的“意图执行”。当任务进入真实工程系统，预定义 endpoint 会过窄，shell 命令又过粗；关键问题变成：Agent 如何承接运行时状态，如何组织临时能力，如何处理异步与反馈，如何把一次探索过程变成可复用的工程资产。

​我们会从 Function Calling、MCP 到 CLI Agent 的演进切入，讨论为什么它们都在指向同一个新的执行层范式。最后，分享将结合我们在 Unity 游戏开发中的实践，展示 REPL 如何进入复杂引擎环境，处理场景、资源、运行时状态与异步流程，并在实际生产任务中展现出远超传统 MCP endpoint 模式的效率与表达能力。

​We believe that the execution layer for next-generation agents will neither stop at MCP nor settle at CLI — it will evolve toward a new paradigm exemplified by the REPL. MCP standardizes tool integration, and CLI unlocks the composability of system environments, but complex engineering environments demand an execution layer that is more continuous, more expressive, and closer to the runtime: a working interface that enables agents to explore, verify, execute, and consolidate capabilities within real systems.

​This talk will develop around that thesis: the agent's execution layer is moving from "API invocation" and "command execution" toward a higher level of abstraction — "intent execution." When tasks enter real engineering systems, predefined endpoints become too narrow, while shell commands become too coarse. The critical questions become: How does an agent take over runtime state? How does it organize ad-hoc capabilities? How does it handle asynchrony and feedback? And how does it turn an exploratory process into reusable engineering assets?

​We will trace the evolution from Function Calling and MCP to CLI Agents, discussing why they all point toward the same emerging execution-layer paradigm. To conclude, we will draw on our experience in Unity game development to demonstrate how a REPL can operate within a complex engine environment — handling scenes, assets, runtime state, and asynchronous workflows — and show how, in real production tasks, it delivers significantly greater efficiency and expressiveness than the traditional MCP endpoint model.

【参与指南】

​请注意：本活动将在 MuShanghai 举办。
除完成本活动的 Luma 报名外，参与者还需要额外持有 MuShanghai 的票以进入场馆。

​你可以根据自己的需求选择以下任一票种：

​Daily Pass 当日票
适合只参加当天活动的朋友。

​Full Pass 月票
适合希望在一个月内使用 MuShanghai 共享办公空间、参加更多活动的朋友。

​购票链接：https://p.baominggongju.com/detail?eid=69dcff11a2b2507f32dc4372

​Host / 主办方：大雄 窦月汐 胡一鸣Daniel
Partner community / 合作社区：游戏客厅