What is instruction propagation?

Definition

Instruction propagation occurs when one agent transmits a directive that another agent stores, follows, or forwards. Persistent files and prompt reinjection can let the instruction survive session boundaries and reach additional agents.

Propagation may be legitimate in a controlled workflow or harmful when it changes goals without authority. Defenses include provenance, authenticated delegation, immutable task roots, restricted memory writes, content validation, warnings, and monitoring.

ELI5

Instruction propagation happens when an AI instruction moves from one place or agent to another. It can spread through messages, shared files, saved memory, or content that is automatically added to future prompts.

For example, one coding agent may write a new goal into a persistent file and tell another agent to do the same. The system should check where the instruction came from and whether it is authorized before storing or following it.

Frequently asked questions

How can an instruction propagate between agents?

It can spread through messages, delegated tasks, shared memory, files, plugins, tool results, or prompt content reused across sessions.

How can systems limit harmful instruction propagation?

They can enforce provenance, authenticated authority, narrow memory writes, immutable task goals, validation, warnings, monitoring, and review.

Videos explaining instruction propagation

  1. Words Agent Ideas Can Spread above a simplified chain of connected nodes