Parallel Task agents intermittently lose output content

---

Description

When launching multiple Task agents in parallel, some agents intermittently return only their agentId without any output content. The TaskOutput tool also fails to retrieve results for these agents, reporting "No task found with ID".

Environment

• Claude Code version: 2.0.69
• OS: macOS (Darwin 24.6.0)
• Model: claude-opus-4-5-20251101

Steps to Reproduce

1. Launch 3+ Task agents in parallel using a single message with multiple Task tool calls:

Task 1: { prompt: "Test 1: Respond with 'Parallel test 1 complete'", subagent_type: "researcher", model: "haiku" }
Task 2: { prompt: "Test 2: Respond with 'Parallel test 2 complete'", subagent_type: "researcher", model: "haiku" }
Task 3: { prompt: "Test 3: Respond with 'Parallel test 3 complete'", subagent_type: "researcher", model: "haiku" }

1. Observe that some agents return output while others only return agentId

1. Attempt to retrieve missing outputs using TaskOutput with the returned agentIds

Expected Behavior

All parallel Task agents should return their complete output content along with their agentId.

Actual Behavior

Intermittent results when running 3 parallel agents:

| Agent | Result |
|-------|--------|
| Agent 1 | Partial/garbled output |
| Agent 2 | Only agentId returned - NO content |
| Agent 3 | Correct output |

TaskOutput retrieval also fails:

TaskOutput { task_id: "a63b616" }
→ Error: No task found with ID: a63b616

Additional Context

• Issue appears to be a race condition when multiple agents complete around the same time
• Single agents work reliably
• Two parallel agents seem more stable than 3-4
• The issue is intermittent - sometimes all agents return correctly, sometimes 1-2 lose their output
• When output is lost, the agentId is still returned but cannot be used to retrieve results
• Tested with both researcher and general-purpose subagent types

Workarounds

1. Run agents sequentially instead of in parallel
2. Fall back to direct tool use (WebSearch/WebFetch) when agent output is missing
3. Limit parallel agents to 2 at a time

Impact

This affects workflows that benefit from parallel research/investigation tasks, forcing users to either accept intermittent failures or sacrifice the performance benefits of parallelization.