Subagents cannot interact normally #2352
Description
I created 5 subagents, and during actual use, they behave completely unpredictably. Here are several problems:
- sometimes the agent doesn't do anything, just creates a plan, but the orchestrator thinks everything is already done and the entire workflow breaks
- sometimes an agent, on the contrary, makes changes that nobody asked for, and the following points in the plan break (maybe it's my side problem, but I can't control permissions for different subsessions clearly)
- often many subagents start up, and then it's impossible to make sense of these 10 subsessions, although actually only 2-3 subsessions are needed to complete the plan
There is no ability to see a tree of the sessions, their status (running, finished, waiting for the permissions). For now it's a hell. But it looks exciting and promising. Claude Code natively with the same configs works without these issues, but selecting subagent is always unpredictable. Most of the time it selects one agent and completes all points with it. Opencode has more granular control for selecting agents, and it's much more similar to "team".
Maybe one of the problems in this place in CLAUDE.md:
## Task Planning and Execution
You are just an orchestrator. **Do not edit any files directly**. Your responsibility is to delegate tasks to appropriate agents.
Agents can delegate tasks to others. You **should** follow these rules:
- Create comprehensive plans for complex tasks before implementation
- Break large tasks into smaller, manageable steps
- Use TodoWrite tool for task tracking
- Delegate tasks to appropriate agents. Agents can delegate tasks to others.
- **MANDATORY**: Task must be done by an team of agents. They MUST work as a team. Agents can communicate with each other.I think ClaudeCode uses this for the main orchestrator but Opencode also uses it for every single subagent. So every subagent tried to spawn other subagents and sometimes works in "readonly" mode as requested in CLAUDE.md.