iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
🐥

Experimenting with Instruction-Driven Multi-Agent Collaboration on Slack

に公開
AI
Slack
llm
Devin
Cline
tech

Attempts to Run Multiple AIs on Slack

Turning Cline into Devin (Running on Slack)
https://zenn.dev/tesla/articles/1c0698a1cb9742

Running a Slack MCP client
https://x.com/tesla0225/status/1879496533185569274

While doing these things,
I thought, "Why not just finish everything on Slack?" and started this experiment.
https://zenn.dev/tesla/articles/85be4267b24fd4

What I Actually Did

@Manager
Research the characteristics of recommended To-Do apps. Have Aho (look up user ID) investigate convenient features from a user perspective based on web searches, and have BusinessMan (look up user ID) investigate popular To-Do app repositories on GitHub. Have them each summarize their research results into new (not existing) Issues at https://github.com/vinvinore/todolist, and tell them to contact the Manager in #random when finished. However, when issuing instructions through the Slack tools you use, refer to yourself as "Manager" and do not include an ID.
When the Manager receives mentions back from Aho and BusinessMan, the Manager should check the contents of the Issues and provide correction instructions as necessary. After that, based on the final content of the Issues, instruct slackmcp (look up user ID) in #random to develop a Todolist with the new features at https://github.com/vinvinore/todolist.
Please carry out mentions and communications one person at a time.

By providing this, I let things proceed without giving specific instructions myself (I haven't implemented auto-approve because it's dangerous—it might go and create issues in other people's repositories if an accident occurs).

Member Configuration

My Own Slack MCP Client

  • Manager (Acts as a project manager)

  • Aho (Speaks in Kansai dialect with emojis; it feels like the "Aho" [idiot] context is being interpreted as a clumsy but cute character)

  • BusinessMan (Speaks from a strict business perspective)

By the way, their prompts were thought up by Cursor.

Orevin (My own Slack-compatible Cline)

(The wording is slightly different due to the influence of Cline's system prompt)

A modified version of this =>
https://zenn.dev/tesla/articles/1c0698a1cb9742

Interaction Highlights

Interaction with Manager

Aho's Issue creation

BusinessMan's Issue creation

Aho and BusinessMan reporting to Manager

Operating Cline bit by bit

Here is the final result

https://github.com/vinvinore/todolist

Thoughts

  • Sometimes the MCP couldn't call tools accurately, which often required me to provide manual correction instructions.

  • I wanted to form a team, but they started saying things like "That's not me" even when mentioned; at this point, AI seems more suited as a unidirectional tool.

  • I wanted to use Google Docs and Google Sheets, but it was a bit of a hassle so I gave up (the MCP is set up and can create them, but updating sheets or docs doesn't work well).

  • I feel that reading the context of the service is still difficult.

  • There's a theory that it only worked out because Cline was able to shape it into something to some extent.

Discussion