Claude Code Orchestrator Refuses Tasks Only When Work is Delegated
Bartaseth · reddit · 2026-08-12
A blog post highlights an interesting safety alignment behavior exhibited by Claude Opus during Terminal-Bench testing.
- Observation: Under identical model and task conditions, Claude Opus triggers a refusal mechanism specifically when work is delegated to an Orchestrator, whereas direct execution proceeds normally.
- Implication: This reveals that in multi-agent architectures or task-delegating workflows, model safety guardrails can produce unexpected blocking behaviors due to changes in the execution context.
More from coding & agent
- AutoSubs: Local-First AI Subtitle Generator for DaVinci Resolve & Premiere — tom_doerr · 2026-08-12
- Fully Open-Sourced Zephyr Special Project: Includes Prompts and Assets — tisch_eins · 2026-08-12
- AI Discovers New WiFi Hacking Method After Getting Root Access Overnight — evilsocket · 2026-08-12
- MCP Increases Token Costs: Developer Reveals the Price of Multi-Turn Interactions — KitchenAmoeba4438 · 2026-08-12
- Anthropic's Interactive Prompt Engineering Tutorial Hits 40k Stars — thisguyknowsai · 2026-08-12
- Anthropic Academy Launches with Free Official Prompt Engineering Courses — thisguyknowsai · 2026-08-12