Serena as planner-researcher subagent?

[myarcana]

I've just read this good article about subagents: https://faafospecialist.substack.com/p/vb-04-subagents-from-basic-to-deep

Key takeaways

• Context Management / Context Engineering is critical!
• Context Rot: LLM performance degrades as input length increases.
• Protect the Main Context at all costs.
• Use the File System as Memory to streamline context.
• Process for building up useful context: “Explore - Planning - Execute.”
• Common Mistake: Using Sub Agents for Implementation.
• Sub Agents are an army of useful information collectors! (Context Collector Army)

[...]

As you can see: None of the agents handle the IMPLEMENTATION itself!

All agents now only serve to collect useful information for the Main Context Window of the primary agent (Claude Code).

Then, they feed the plans/reports back to the primary agent (Claude Code itself) to begin execution

I feel it would be very useful to have a subagent that uses Serena to explore the project and explain things about it to Claude.

How could I set that up?

One problem (?) I noticed with running Claude Code with Serena MCP in the first place

claude mcp add serena -- \
uvx --from git+https://github.com/oraios/serena serena \
start-mcp-server --context ide-assistant --project "$(pwd)"

in the first place is that after claude code started up, claude was active and interactable for quite a long time before serena finally started, leading to a conversation like this

**> Are you using Serena MCP? **

⏺ No, I'm not currently using Serena MCP. I don't see any MCP (Model Context Protocol) servers connected in this session.

  MCP-provided tools would appear with an "mcp__" prefix in my available tools. If you'd like to use Serena MCP or any other MCP server with Claude Code, you would need to configure
  it in your settings.

  Are you looking to set up an MCP server, or were you checking for a specific functionality that Serena MCP provides?

**> What about now? Check again **

⏺ You're right! Yes, I do have access to Serena MCP. Looking at my available tools, I can see many tools with the **mcp__serena__** prefix, including:

  - Symbol-based code navigation and editing tools (**find_symbol**, **replace_symbol_body**, **insert_after_symbol**, etc.)
  - Semantic search capabilities (**search_for_pattern**, **find_referencing_symbols**)
  - Code overview tools (**get_symbols_overview**)
  - Memory management for storing project information
  - Project management tools

  These tools are designed for intelligent, token-efficient code exploration and editing by working with code symbols (classes, methods, functions) rather than just raw text. This
  allows me to navigate and modify codebases more precisely.

  Sorry for the confusion in my initial response - Serena MCP is indeed active and available for this session!

I worry that in a naive set up of an subagent with serena-via-mcp, the subagent would start getting to work before serena is ready

I've seen everywhere here it's recommended to start Serena manually in SSE mode. Is there a way I can instead make it so that when a subagent needs serena, if serena is not running, it starts it itself? I wouldn't want to have serena always running in the background because I don't always need it.

[MischaPanch]

This is really a discussion for the client side of things. Claude Code behaves like this (not waiting for MCP servers to start) since rather recently, other clients don't behave like that at all. You can easily turn off the SSE server through the dashboard, so you don't need to keep Serena running always.

One thing that occurs to me would be to release Serena as an app (in addition to the standard release as python package now). This is something that the community could do, or we may get to it later (not a priority now)