Agent tool silently rejects first agent in parallel group without showing permission prompt

Description

When launching multiple agents in parallel (multiple Agent tool calls in a single message), the first agent in the group is silently rejected without ever displaying a permission prompt to the user. The remaining agents in the same group launch and execute normally.

Reproduction

1. Use Claude Code CLI (v2.1.80) with default permission settings
2. Send a message that triggers multiple parallel Agent tool calls (e.g., 5 agents launched simultaneously)
3. The first agent in the group is rejected with error "The user doesn't want to proceed with this tool use"
4. No permission prompt is ever shown to the user for that first agent
5. The remaining 4 agents display their permission prompts normally and execute successfully

Frequency

This has been reproduced 3 times in the same session:

• First occurrence: 5 agents launched in parallel, agent 1 silently rejected, agents 2-5 succeeded
• Second occurrence: Same agent re-launched solo, silently rejected again (possibly related to the first failure's state)
• Third occurrence: Same agent re-launched solo again, silently rejected again

Environment

• Claude Code version: 2.1.80
• Model: claude-opus-4-6 (1M context)
• Platform: macOS (Darwin 25.3.0)
• Shell: zsh

Expected Behavior

All agents in a parallel group should display their permission prompts to the user for approval/denial.

Actual Behavior

The first agent is silently rejected. The user sees no prompt and has no opportunity to approve or deny. The error message incorrectly states "The user doesn't want to proceed" when the user was never asked.

Additional Context

A separate but possibly related issue was observed with background agents (run_in_background: true): Write tool permission prompts from background agents are also silently denied without surfacing to the user. This may share a root cause — permission prompts that can't be displayed to the user are auto-denied rather than queued.