claude-code - 💡(How to fix) Fix Persistent 529 overloaded errors causing subagent failures mid-execution [2 comments, 2 participants]

• Multiple subagents launched in parallel
• Agents accumulate 40-90+ tool uses of progress, then hit 529 and terminate
• Retrying immediately fails again (529 within seconds, 0 tool uses)
• All work done by the agent is lost since it cannot resume from where it stopped
• Status page showed "All Systems Operational" during the incident

TL;DR

Implementing retry logic for transient 529 errors or preserving partial results could mitigate the issue of subagents terminating mid-execution.

Guidance

• Investigate the Agent/Task tool's configuration to see if there are any settings that can be adjusted to enable retry logic for 529 errors.
• Consider modifying the subagent execution workflow to save intermediate progress at regular intervals, allowing for resumption from the last saved point in case of termination.
• Review the Claude Code CLI documentation to determine if there are any built-in mechanisms for handling or retrying on transient errors like 529.
• Examine the Opus model's behavior to understand if it has any influence on the subagents' termination due to 529 errors.

Example

No specific code snippet can be provided without more details on the Agent/Task tool's API or the Claude Code CLI's error handling mechanisms.

Notes

The solution may require modifications to the Agent/Task tool, the Claude Code CLI, or the workflow design, depending on where the flexibility for retry logic or progress preservation can be introduced.

Recommendation

Apply workaround: Implementing a retry mechanism or preserving partial results seems more feasible given the current information, as it does not require waiting for a potential fix in the underlying tools or models.