Broken Telephone in Agent Chains - Why Intent Gets Lost Beyond 2 Hops
Broken Telephone in Agent Chains
You ask Agent A to "update the pricing page to reflect the new enterprise tier." Agent A delegates to Agent B: "modify the pricing component." Agent B delegates to Agent C: "change some values in the pricing file." By the time Agent C acts, the original intent is gone. It changes random values because "change some values" is all it knows.
This is the broken telephone problem in agent chains, and it gets worse with every hop.
Intent Loss Is Exponential
Each handoff between agents loses context. The first hop usually preserves the core meaning. The second hop starts dropping important details. By the third hop, the executing agent often has no idea why it is doing what it is doing.
This matters because without intent, agents cannot make judgment calls. Should the new tier be highlighted? Should it appear first or last? Agent C has no way to answer these questions because it never received the business context.
The Central Coordinator Pattern
The fix is architectural: never chain agents linearly. Use a central coordinator that holds the original intent and delegates directly to specialized agents.
Instead of A -> B -> C, the pattern is:
- Coordinator holds the full context and goal
- Coordinator -> Agent 1 gets a specific, well-scoped task with full context
- Coordinator -> Agent 2 gets a different task, also with full context
- Results flow back to the coordinator, which assembles the final output
Each agent is one hop from the original intent. No telephone game.
Practical Implementation
The coordinator does not need to be complex. It can be a single prompt that breaks a task into sub-tasks and dispatches them. The key constraint is that every sub-task includes the original goal, not just the immediate action.
Keep your agent chains flat. Two hops maximum. Your output quality will improve dramatically.
Fazm is an open source macOS AI agent. Open source on GitHub.