Persistent 529 overloaded errors causing subagent failures mid-execution

When using Claude Code with background subagents (via the Agent/Task tool), 529 overloaded errors are terminating agents mid-execution with no recovery path.

Observed behavior

• 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

Impact

• Subagents that run for 5-15 minutes and accumulate significant progress are killed without warning
• No graceful degradation or retry logic at the subagent level
• Users must manually relaunch agents from scratch, losing all intermediate progress
• Effectively blocks workflows that depend on parallel agent execution

Expected behavior

• Subagents should retry on transient 529 errors before terminating
• Or at minimum, partial results should be preserved so agents can resume

Environment

• Claude Code CLI
• macOS
• Opus model