Lowe•2mo ago

Excess Responses from Claude 4.5 Haiku

Hi everyone, I’m currently building a multi-agent workflow using one primary agent (based on Claude 4.5 Haiku) that serves as both the main assistant and a router when needed. The goal is for this main agent to handle user interactions within specific boundaries. When the conversation shifts topics or falls outside its scope, it should silently route the query to another agent — without sending any message or executing an exit condition first. However, despite my prompt explicitly forbidding any message or exit action in these cases, the agent still always sends a message and triggers an exit. I’ve experimented with several prompt variations and structural changes, but none have worked so far. Switching LLMs isn’t an option right now since Claude is the only model that meets our quality requirements — and Haiku is significantly more cost-effective than Sonnet. Has anyone dealt with similar behavior or found a reliable workaround for this routing issue? Images: - 1: Shows current prompt for handling routing (multiple different approaches, guidelines, etc. has been tried. Also, a seperate "restrictions" section in the prompt also prohibits this.) - 2: The highlighted message shows where the AI is ignoring our restrictions (sending message and responds in Swedish instead of English. Translated: "That's a good question! 🚚 I'll put you through to our support team who can give you exact delivery times and answer any logistics questions.") Any insights would be greatly appreciated. — Lowe