529 Overloaded errors kill parallel subagents, no retry — work lost

Problem

When running multiple subagents in parallel (a core Claude Code workflow), API 529 "Overloaded" errors silently kill agents with no automatic retry. The agent reports completion but the result is just an error string. All work context is lost.

What happened

I was running 3 parallel agents (SEO metadata, Performance audit, Accessibility audit) for a Next.js e-commerce project. All 3 hit 529 errors:

API Error: 529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}

• Agent 1 (SEO): still running (survived somehow)
• Agent 2 (Performance): dead — 47 tool uses attempted, 0 useful output
• Agent 3 (Accessibility): dead — 51 tool uses attempted, 0 useful output

Each agent had consumed tokens and time, but returned nothing. I had to manually re-dispatch them.

Impact

• Lost work: Each agent had already read files and started analysis before dying
• Lost tokens: ~4000 tokens consumed with zero output
• Lost time: ~10 minutes per agent × 2 agents = 20 minutes wasted
• User frustration: This is the second time in this session. Parallel agents are a flagship workflow — they should be reliable.

Expected behavior

1. Automatic retry with backoff — 529 is transient. Wait 10-30 seconds and retry, don't just die.
2. Partial result preservation — If the agent did useful work before the error, save and return what it had.
3. Clear notification — Tell the user "Agent X failed due to API overload, retrying..." instead of silently completing with an error blob.
4. Graceful degradation — If 3 parallel agents all hit 529, maybe serialize them instead of failing all 3.

Environment

• Claude Code CLI on Windows 11
• Model: claude-opus-4-6 (1M context)
• 3 parallel agents via Agent tool with run_in_background: true
• Session had been running ~3 hours with heavy parallel agent usage (15+ agents total)

Reproduction

1. Start a session with heavy context (~500K tokens)
2. Launch 3+ parallel agents with substantive tasks (file reads + writes)
3. Wait — during peak hours, 529 errors are common
4. Observe: agents die silently, status: completed but result is just the error

Suggested fix priority

This is high priority because parallel agents are the main way Claude Code delivers complex multi-file work. If they're unreliable, users lose trust and stop using the feature.

---

Submitted from an active Claude Code session by a frustrated but loyal user.