openclaw - 💡(How to fix) Fix Per-parent concurrency cap on subagent spawns [1 comments, 2 participants]

TL;DR

Implementing a configurable per-parent concurrency limit can help mitigate the event loop saturation caused by multi-wave fanout patterns.

Guidance

• Introduce a concurrency limit (e.g., 2-3) for subagent spawns to prevent overwhelming the event loop.
• Queue subagent spawns that exceed the concurrency limit to prevent contention for the WebSocket.
• Monitor CPU usage and Slack pong timeouts to verify the effectiveness of the concurrency limit.
• Consider adjusting the default concurrency limit based on the specific requirements of the agent fleets.

Example

# Pseudo-code example of queuing subagent spawns
class Agent:
    def __init__(self, concurrency_limit):
        self.concurrency_limit = concurrency_limit
        self.spawn_queue = []

    def spawn_subagent(self):
        if len(self.spawn_queue) < self.concurrency_limit:
            # Spawn subagent immediately
            self._spawn_subagent()
        else:
            # Queue subagent spawn
            self.spawn_queue.append(self._spawn_subagent)

    def _spawn_subagent(self):
        # Actual subagent spawn logic
        pass

Notes

The proposed solution assumes that the event loop saturation is primarily caused by the contention for the WebSocket among subagents. However, other factors may contribute to the issue, and additional investigation may be necessary to fully resolve the problem.

Recommendation

Apply a workaround by introducing a configurable per-parent concurrency limit, as this can help mitigate the event loop saturation without requiring significant changes to the existing codebase.