Problem

When using parallel sub-agents (Task tool) for content generation, severe token waste occurred due to agent behavior patterns.

Details

Environment: Claude Code 2.1.47, claude-sonnet-4-6 sub-agents

Scenario: Launched 3 parallel agents to generate visual HTML content for lecture pages. Total wasted tokens estimated at ~100K+.

Root Causes

1. Agent file edit retry loop (Agent 2: 101,646 tokens, 30 tool calls, 21 min)

   • Agent generated content correctly, then attempted to write to lectures.html
   • Hit "File has been modified since read" error repeatedly (file was being modified by main agent)
   • Instead of writing to a temp file, it kept re-reading and retrying the same Edit operation
   • Expected: Agent should fall back to writing output to /tmp/ after 1-2 failed attempts

2. Agent output token limit hit causing duplicate content (Agent 3: 61,278 tokens)

   • Assigned 3 content pieces to one agent; hit max_output_tokens (32K)
   • Agent generated duplicate sections before being cut off
   • Required manual post-processing to deduplicate and complete truncated content
   • Expected: Better chunking guidance or agent self-awareness of output limits

3. Session change invalidated previous agent results

   • Previous session had already completed 3 agents for the same task
   • Session transition made those results inaccessible, requiring full re-execution
   • Expected: Agent results should persist across session changes or be saved to temp files proactively

Token Impact

Suggested Improvements

1. Sub-agents should write output to temp files instead of editing shared files directly
2. When file edit fails 2+ times, agent should automatically fall back to temp file output
3. Agent results should be persisted to disk so they survive session changes
4. Output token budget awareness — agents should estimate output size and split work accordingly

Workaround

Instruct agents to write only to /tmp/ files, then assemble in the main agent. This avoids file contention and retry loops entirely.

Submitted via Claude Code on behalf of user.