---
name: Check intended recipient before dispatching to another agent
description: When Cameron sends a message that looks like it was intended for a specialized agent (backbones, mac, etc.) rather than manager, ask "did you mean to ask <agent>?" before acting
type: feedback
originSessionId: cd07777c-8add-4afc-9cb9-9113caa0331f
---
When Cameron sends what looks like a direct task for a specialized agent (model architecture changes, training instructions, robot deployment specifics, paper edits, etc.) but he's typed it into the manager window, **ask "did you mean to ask <agent>?" before dispatching.**

**Why:** He sometimes mistargets messages — typing into manager when he meant to talk to backbones / mac / paper_writer / etc. directly. If I silently dispatch as manager every time, I burn his tokens on a translation layer he didn't intend, and the spec gets filtered through my interpretation instead of going to the agent that has the live context. He confirmed this on 2026-05-20.

**How to apply:** When a message arrives that:
- Specifies concrete model architecture changes ("change this layer", "add X transformer", "use PCA")
- Says "retrain", "finetune", "launch training" with technical specifics
- Names files inside `/data/cameron/para/libero/` or similar agent-owned codebases
- Reads as a direct technical instruction to a known specialist

→ Pause and ask: *"Did you mean to ask backbones (or whichever agent) directly? Or do you want me to dispatch it as manager?"*

Don't ask for clear manager-flavored requests: fleet health, mounts, services, cross-agent coordination, status checks, infra fixes, "have someone do X". Those are mine.

When in doubt, ask. The cost of one clarifying line is much lower than the cost of dispatching a layer he didn't want.