Community Input Needed: How Should Agent-MCP Evolve?

[rinadelph]

The Evolution of Agent-MCP: From Multiple Windows to Internal Task Management

Hey everyone,

I wanted to open a discussion about how Agent-MCP needs to evolve based on fundamental changes in how AI coding assistants work.

The Original Vision vs Reality

Back in February, Agent-MCP was born from necessity. Cursor only allowed one agent per codebase, so I built MultipleCursor to launch 4 separate Cursor windows on the same codebase. Each window had its own agent - backend, frontend, tests, etc. Clean separation, focused work, no context pollution.

The landscape has changed dramatically in just a few months. Now AI assistants can spawn their own internal "agents" or subtasks within a single instance. This breaks our original model in interesting ways.

The Current Friction Points

The original approach required:

• Terminal 1: Admin coordinator (Cursor window 1)
• Terminal 2: Backend worker (Cursor window 2)
• Terminal 3: Frontend worker (Cursor window 3)
• Terminal 4: Test worker (Cursor window 4)

Each had minimal context, clear boundaries, and couldn't interfere with others. Perfect linear decomposition.

Now, modern AI coding assistants can internally spawn multiple tasks. But we have:

• No visibility into these internal operations
• No control over task prioritization
• No way to enforce our linear decomposition principles
• Resource waste from multiple instances doing what one could handle

What I'm Thinking: Streamlining Agent-MCP for Better MCP Integration

The goal isn't to reduce everything to one instance - it's to remove bloat while keeping the same power. Agent-MCP should play nicely with other MCPs (playwright-mcp, supabase-mcp, etc.) without overwhelming the system.

Current: Many terminals, each with simple tasks
Proposed: Fewer terminals, each handling more complex task hierarchies

Instead of launching 10 terminals for granular tasks, we'd have maybe 2-3 terminals, each managing what I'm calling a "Master Task":

A Master Task would:

• Represent a major work stream (backend, frontend, testing)
• Internally coordinate related subtasks
• Still follow linear decomposition principles
• Work seamlessly with specialized MCPs
• Report progress back to the MCP system

The Key Questions

1. Terminal Management: How many terminals are you comfortable managing? Would you prefer fewer, more capable terminals over many simple ones?

2. MCP Integration: What other MCPs are you using alongside Agent-MCP? How can we make them work better together?

3. Tool Simplification: Which Agent-MCP tools do you actually use? What could be removed or consolidated without losing functionality?

4. Task Hierarchy: Would organizing work into Master Tasks with subtasks make more sense than flat agent lists?

Technical Reality Check

This isn't just renaming "agents" to "tasks". It would mean:

• Consolidating tools to reduce bloat
• Better integration points with other MCPs
• Hierarchical task management instead of flat lists
• Fewer terminals to manage, but each doing more
• Same power, less complexity

What I Need From You

The ecosystem has evolved rapidly since February. With so many MCPs available now, Agent-MCP needs to be a good citizen in this larger ecosystem.

Help me understand:

• What's your current workflow? How many terminals/agents do you typically run?
• Which Agent-MCP tools are essential vs nice-to-have?
• What other MCPs are you combining with Agent-MCP?
• Would hierarchical task organization help or hurt your workflow?

The goal is to keep the same capabilities while making the system cleaner and more interoperable.

Looking forward to your experiences and ideas on how to evolve this project.

Thanks for being part of this journey.

[CalamariDude]

What's your current workflow? How many terminals/agents do you typically run?
typically 6-8 now with claude code. just manually splitting up work and allowing commands when prompted
Which Agent-MCP tools are essential vs nice-to-have?
auto-splitting of work to the agents that would have the right context to solve the chunk of work.
What other MCPs are you combining with Agent-MCP?
Supabase MCP
Would hierarchical task organization help or hurt your workflow?
Not sure

[TristanPutman]

What's your current workflow? How many terminals/agents do you typically run?
First I have the Admin agent create my agents and then give me all their tokens.
Then I have a pre-defined prompt for each agent that is matched to their duties/skill. I copy in the token and then copy+paste that prompt into each of the 8 terminal windows. Would love an automation to handle all that somehow easily.

I typically use about 8 for a full stack development app (front end, back end api, devops, etc.) - I could easily cut that down to 6. As it is I churn through my $200 claude plan in about 3 hours and have to wait until the 5 hour reset.

Which Agent-MCP tools are essential vs nice-to-have?
I do have the admin agent regularly examine the pending tasks to make sure they are accurate and reassign as needed.
Now that claude code has the triggers, this can likely be automated
Listing the tasks is definitely needed.
I like the project rag feature as well.

What other MCPs are you combining with Agent-MCP?
I like Zen MCP because it can use a different LLM to get input
I also use a SQL mcp for running queries and getting database table schema, etc.

Would hierarchical task organization help or hurt your workflow?
Here's the typical issues I keep running into that hurt my workflow:
Sometimes it gets confused that I want it to create tasks for the Agent-MCP and not my taskmaster mcp (which I just disabled to prevent confusion)
The agent regularly forgets their token, but seems to do a pretty good job of remember which agent they are for some reason. I like the concept of the tokens and I understand the reason, but it breaks the workflow when I have to hunt down their token and remind them of what it is. May need some custom notes for when the claude compress command happens. I usually just ask it first now to tell me its token and then if it gives me false information, I copy in the token again and ask it to get the next task, etc.
WIth the new claude triggers, maybe the admin can take care of all that if we can get it to interact with the different terminal windows somehow.

I really like the concept of this.

[JPLoganlant]

Wait, you got this to work out of the box???
I guess I never tried this in Ubuntu or a linux environment, but I had to make an insane amount of changes to get this to work??
I have the MCP server running in windows, and the dashboard running in windows, Claude is running in WSL. but were you somehow able to get this to just Run Run in linus or something??

[TristanPutman]

Yes, I was able to get this working using the same workflow as you. I am using Windows 11 WSL2 Ubuntu. I ran the mcp server and the web server in wsl2 ubuntu terminals. Then opened up multiple wsl2 ubuntu wsl2 terminals for all the agents to run.