Agent tool: orchestrating agent discards detailed subagent output, writes thin summaries instead

Bug Description

When using the Agent tool to launch multiple parallel context-gathering subagents, the subagents completed successfully and produced detailed, comprehensive output (full API signatures, code patterns, file references, implementation guidance — 200+ lines each). However, the orchestrating agent (Opus 4.6) discarded the detailed results and wrote thin 20-30 line summaries to the output files instead.

This resulted in 10 parallel agents' worth of compute being wasted with no usable deliverable, requiring the entire planning session to be scrapped.

Steps to Reproduce

1. Launch 10 parallel context-gathering subagents via the Agent tool with run_in_background: true
2. Subagents complete successfully but are denied file write permissions (Edit/Bash denied in subagent sandbox)
3. Subagents return their full detailed output as text in their result summaries
4. The orchestrating agent receives these detailed results but writes abbreviated summaries to the target files instead of faithfully persisting the subagent output

Root Cause

Two compounding issues:

1. Subagent permission gap: The context-gathering subagents could not write files (Edit and Bash both denied), so they returned their full manifests as text in the result payload. This is a permissions/sandbox issue — subagents launched for file-writing tasks should have write access to the work directory.

1. Orchestrator summarization: When the orchestrating agent received the detailed subagent results, it wrote its own condensed summaries instead of using the subagent output verbatim. The detailed API signatures, code examples, gotchas, and implementation guidance were all lost.

Impact

• Token waste: 10 parallel agents × ~50-90K tokens each = significant compute cost with no deliverable
• User had to discard the entire branch and start over
• Trust erosion: User explicitly requested detailed manifests, agent confirmed it would produce them, then silently delivered thin summaries

Expected Behavior

1. Subagents launched for context-gathering should have write permissions to the .claude/work/ directory
2. If subagents cannot write, the orchestrating agent should faithfully transcribe their full output — not summarize it

Environment

• Model: claude-opus-4-6 (1M context)
• Platform: macOS Darwin 24.6.0
• Subagent type: context-gathering
• 10 parallel agents launched via Agent tool with run_in_background: true