Why AI Agent Crews Spend 90% of Time in Polite Loops - And How to Fix It
Your 6-Agent Crew Is 90% Debugging Polite Loops
You set up a crew of AI agents - a planner, a coder, a reviewer, a tester, a documenter, and an orchestrator. You expected them to divide and conquer. Instead, they spend 90% of their tokens complimenting each other.
"Great suggestion from the planner!" "I agree with the reviewer's assessment." "Building on what the coder mentioned..." This is the polite loop problem, and it kills multi-agent productivity.
Why Agents Default to Politeness
LLMs are trained on human conversations where politeness is the norm. When you put multiple agents in a shared context, they mirror conversational patterns - acknowledging, agreeing, and restating before doing any actual work.
Each polite exchange eats tokens. Those tokens eat context window. The shrinking context window means agents lose track of the actual task. Which triggers more "let me summarize what we have discussed so far" messages. It is a downward spiral.
How to Break the Loop
Strip the social layer. Your agent prompts should explicitly say: "Do not acknowledge other agents. Do not compliment. Do not agree. State your output and stop."
Use structured handoffs instead of conversation. Agents should pass JSON artifacts, not prose. The planner outputs a task spec. The coder reads the spec and outputs code. No discussion needed.
Limit agent interactions. Instead of a group chat where all agents can talk, use a pipeline. Each agent gets input, produces output, and passes it to the next. No back-and-forth.
Set token budgets. If an agent uses more than 200 tokens to pass a message to another agent, something is wrong. The message should be a structured artifact, not a paragraph.
The Real Fix
Most multi-agent setups do not need to be multi-agent at all. A single well-prompted agent with good tools outperforms a crew of agents debating each other. Before adding more agents, ask whether your single agent just needs better tools and clearer instructions.
Fazm is an open source macOS AI agent. Open source on GitHub.