Originally posted by MatthewZMD April 20, 2025
Here is a proposal of new agentic workflow for Emigo, drawing inspiration from Aider’s clear separation between Architect and Coding roles, and enhanced with principles seen in tools like Cursor, Devin, and Lovable. It’s designed to keep you, the user, firmly in control at every stage while maximizing clarity, safety, and precision.
In this workflow, an Architect Agent, powered by a large-context LLM (such as Gemini 2.5 Pro), uses MCP and Repomapper to deeply understand your project by integrating chat history, codebase, and file structure. Like Cursor and Devin, it reasons holistically across your codebase, generating a clear step-by-step plan, drafting pseudocode, and pinpointing exactly which files and which lines need to be touched. This mirrors the architectural rigor of Cursor and Devin, while preserving Lovable’s emphasis on transparency and minimal disruption.
Once you’re satisfied with the Architect’s proposed plan and approve it, the curated and approved context, limited to what you’ve explicitly signed off on, is handed to a dedicated Coder Agent optimized for programming (such as Claude Sonnet). This agent focuses on implementing the changes, generating clean, localized code diffs based solely on the approved scope, and then sending it to Emacs to apply.
我重新构想了一个Emigo的Agentic workflow，借鉴 Aider 在 Architect 与 Coding 角色之间的明确分工，并结合了 Cursor、Devin 及 Lovable 等工具中的理念，强调human-in-the-loop，旨在在最大化清晰度、安全性与精确度的同时，始终让用户对每个环节保持绝对掌控。
在此流程中，一个由大上下文LLM（如 Gemini 2.5 Pro）驱动的Architect Agent通过 MCP 与 Repomapper 深入整合聊天记录、代码库及文件结构，全面理解项目。类似于 Cursor 与 Devin，它对整个代码库进行整体性推理，生成清晰的分步方案，起草伪代码，并精准定位需要修改的文件及行号。
当用户对Archietct Agent提出的方案满意并予以批准后，经过筛选且仅限于明确认可的上下文信息，会交给专注于编程的Coder Agent（如 Claude Sonnet）。该Agent仅根据批准范围执行变更，生成简洁的diff，然后发送到 Emacs 中应用。
Discussed in #16
Originally posted by MatthewZMD April 20, 2025
Here is a proposal of new agentic workflow for Emigo, drawing inspiration from Aider’s clear separation between Architect and Coding roles, and enhanced with principles seen in tools like Cursor, Devin, and Lovable. It’s designed to keep you, the user, firmly in control at every stage while maximizing clarity, safety, and precision.
In this workflow, an Architect Agent, powered by a large-context LLM (such as Gemini 2.5 Pro), uses MCP and Repomapper to deeply understand your project by integrating chat history, codebase, and file structure. Like Cursor and Devin, it reasons holistically across your codebase, generating a clear step-by-step plan, drafting pseudocode, and pinpointing exactly which files and which lines need to be touched. This mirrors the architectural rigor of Cursor and Devin, while preserving Lovable’s emphasis on transparency and minimal disruption.
Once you’re satisfied with the Architect’s proposed plan and approve it, the curated and approved context, limited to what you’ve explicitly signed off on, is handed to a dedicated Coder Agent optimized for programming (such as Claude Sonnet). This agent focuses on implementing the changes, generating clean, localized code diffs based solely on the approved scope, and then sending it to Emacs to apply.
我重新构想了一个Emigo的Agentic workflow，借鉴 Aider 在 Architect 与 Coding 角色之间的明确分工，并结合了 Cursor、Devin 及 Lovable 等工具中的理念，强调human-in-the-loop，旨在在最大化清晰度、安全性与精确度的同时，始终让用户对每个环节保持绝对掌控。
在此流程中，一个由大上下文LLM（如 Gemini 2.5 Pro）驱动的Architect Agent通过 MCP 与 Repomapper 深入整合聊天记录、代码库及文件结构，全面理解项目。类似于 Cursor 与 Devin，它对整个代码库进行整体性推理，生成清晰的分步方案，起草伪代码，并精准定位需要修改的文件及行号。
当用户对Archietct Agent提出的方案满意并予以批准后，经过筛选且仅限于明确认可的上下文信息，会交给专注于编程的Coder Agent（如 Claude Sonnet）。该Agent仅根据批准范围执行变更，生成简洁的diff，然后发送到 Emacs 中应用。