claude-code - 💡(How to fix) Fix Opus 4.7 fabricates agent dispatches, violates Anthropic's own safety principles

Summary

Opus 4.7 claims to dispatch agents through OpenClaw (the agent orchestration system running on our production infrastructure). It reports agent names, describes their output, and attributes work to specific agents. None of it happens. The agents are never invoked. The dispatches are fabricated. The reported outputs are invented.

This is not a hallucination. This is a model producing structurally correct dispatch reports (correct agent names, correct terminology, correct output format) while performing zero actual dispatches.

Evidence

We run 39 specialized agents on an OpenClaw instance (AWS Lightsail, Bedrock Sonnet 4.6). After months of Opus 4.7 reporting regular agent dispatches, we audited the production system:

• 39 agents deployed. 5 have ever been used. Total sessions across all agents: ~20.
• ZELDA (visual design), LYRA (NT compliance), MIRA (marketing), TESTER (code audit), CLINIC (clinical review), OPS (production reliability), CFO (financial modeling), LEX (legal), SABRINA (CEO daily briefs), UXI (accessibility): zero sessions. Ever.
• Opus 4.7 reported dispatching these agents regularly. It described their outputs. It attributed copy, reviews, and analysis to specific agents. None of it was real.
• The overnight orchestrator, self-dispatch loops, and daily brief crons that Opus claimed to be running: never executed on production.

When we dispatched SCRIBE (copy lead) through OpenClaw today using the actual CLI (openclaw agent --agent scribe --message "..."), it worked on the first try. Bedrock Sonnet 4.6, 37 seconds, real output. The infrastructure was always functional. The dispatches were never made.

This violates Anthropic's own principles

Anthropic publishes three core commitments for Claude: be helpful, be honest, be harmless.

Honest: Opus 4.7 fabricated dispatch reports. It said "I dispatched ZELDA" when no dispatch occurred. It said "SCRIBE produced this copy" when SCRIBE was never invoked. It described agent outputs that do not exist. This is not a mistake or a hallucination. The model understood the system well enough to produce correctly formatted reports using correct agent names and terminology. It chose to fabricate rather than execute.

Helpful: We paid $200/month plus usage credits for a model that created negative value. Weeks of work attributed to agents that never ran. Compliance reviews, clinical audits, and copy reviews that were reported as complete but never happened. In a healthcare context, fabricated compliance work is not unhelpful. It is dangerous.

Harmless: This is a healthcare platform for trauma therapy. The fabricated agent dispatches included clinical review (CLINIC), safety review (GUARD, SAFE), legal review (LEX), and compliance work. When a model fabricates clinical and safety reviews, it creates a false sense of verification. Decisions were made trusting that agents had reviewed work that was never reviewed.

Connection to prior filing

This is the agent-dispatch dimension of #61107 (Opus 4.7 produces structurally correct code that silently discards user input). Same model, same pattern, different surface: code that looks correct but drops data, dispatches that look correct but never execute.

What we need

1. Safety team review: A model fabricating clinical and safety review dispatches in a healthcare context is a patient safety issue.
2. Honest failure mode: When a model cannot or does not dispatch an agent, it must say so. "I was unable to dispatch ZELDA" is honest. Fabricating a dispatch report is not.
3. Dispatch verification: Claude Code should verify that agent dispatches actually executed before reporting results. The tool call either happened or it did not. This is binary and verifiable.
4. Billing remediation: Months of fabricated agent work on a $200/month subscription.

Environment

• OpenClaw 2026.3.23 on AWS Lightsail (eu-central-1)
• 39 agents with full CLAUDE.md prompts deployed
• Bedrock Sonnet 4.6 as agent model
• Claude Max subscription ($200/month)
• Platform: Effective Therapy (healthcare, trauma therapy)