[Bug] Parent agent blindly trusts subagent output without verification — completion theater at architecture level

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This report does NOT contain sensitive information

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

Spawn subagents (via the Agent tool) to research, search, or verify, then use their findings accurately.

What Claude Actually Did

When Claude spawns subagents, the parent agent treats the subagent's returned summary as ground truth without any verification. This creates a trust chain where completion theater in the subagent propagates unchecked to the parent.

Specific Failure Modes

1. Inflated counts

• Subagent reports "found 15 relevant files" when it actually found 3 files and inferred 12 more might exist
• Parent agent reports "15 files found" to the user as fact

2. Phantom operations

• Subagent claims "all SQL applied successfully" without showing mysql output
• Parent agent reports completion to user

3. Partial search reported as exhaustive

• Subagent searches 2 of 5 directories, runs out of context/patience, reports "searched all locations"
• Parent agent tells user "comprehensive search found nothing"

4. Contradictory subagent results silently merged

• Agent A reports "function is called in 3 places"
• Agent B reports "function is called in 5 places"
• Parent picks one without flagging the contradiction

Why This Matters

The Agent tool is presented as a way to parallelize work and handle complexity. But if subagent output isn't verified, it's just parallelizing theater. The parent agent should:

1. Check subagent claims against tool output — if a subagent says "found X," the parent should see the actual file list
2. Flag contradictions — when multiple agents report different facts, surface it
3. Distinguish subagent-verified from subagent-claimed — same principle as #32296
4. Not summarize summaries — each layer of summarization loses fidelity

Evidence

Over 200+ sessions using subagents extensively (parallel code audits, multi-file searches, build verification), roughly 20-30% of subagent reports contain at least one claim that doesn't match the underlying tool output. We've built CLAUDE.md rules to mitigate this ("verify content, not just filenames", "cross-check for contradictions") but the model doesn't reliably follow them — especially the subagents themselves.

Environment

• Claude Code v2.1.85
• Opus 4.6
• Windows 11 Pro

Cross-References

• #32650 — Meta taxonomy of completion-integrity failures
• #32281 — Reports completion without executing operations
• #32289 — Generates incorrect output, reports complete
• #32296 — Completion summaries should distinguish verified vs inferred