[Question]: Optimize Agent Loop for Stable Long Automatic Task Execution

## Question or topic

I would like to put forward an optimization suggestion for QwenPaw agent loop execution logic, to improve stability and user experience for long and multi-step automatic tasks.

## Context

I am using QwenPaw for daily automated workflows, including document processing, code execution and multi-skill chained tasks.
Currently the agent loop has several usability limitations in long automatic scenarios:

1. The main agent and all sub-agents share a single global operation limit. Child tasks can easily use up all available steps and cause the entire task to stop halfway.
2. When reaching the maximum operation limit, the task is terminated directly without summarizing results, requiring users to manually continue the conversation.
3. Fully automatic long-running multi-step tasks only work in Mission mode. Ordinary chat sessions cannot run continuous automation stably.
4. Multiple parallel sessions will compete for execution quota, resulting in task interruption and client lag.

## Tried so far

I have compared the local implementation of QwenPaw ReAct loop with Hermes-Agent’s mature conversation loop design.
I have tried increasing the global max iterations and relying solely on Mission mode, but these are only temporary workarounds and cannot fundamentally solve the resource competition and abrupt termination problem.

## Suggested Optimization Behavior

I hope the agent loop can be optimized incrementally without changing the existing ReAct core framework, with the following improvements:

1. Independent execution limits for main agent and sub-agents. Sub-task execution will no longer consume the main task’s execution opportunities, avoiding premature task termination.
2. Add a graceful ending mechanism. When execution steps are about to run out, the agent will automatically summarize existing results before stopping, instead of breaking off suddenly. Simple code execution can be exempted from step consumption to improve utilization.
3. Support native long-task automatic execution in ordinary chat mode. No need to enable Mission mode for continuous multi-step automation.
4. Optimize long conversation stability with automatic context compression to reduce lag and failure rate for long-running tasks.

---