If the AI 'boss' coordinating a team of AI agents makes one bad call, does every agent beneath it just follow along?

This explores what happens when the coordinating agent at the top of a multi-agent system makes a mistake, and how that mistake moves through the agents below it. None of these notes studies orchestrator errors on their own. Taken together, though, they show something useful: a hierarchy rarely contains a top-level error. It tends to amplify the error, hide it, and pass the cost to someone else.

The spreading part is the easiest to picture. Researchers describe four ways a failure crosses from one agent to another. Messages carry influence downstream. Shared state keeps a bad assumption alive over time. Aggregation merges outputs that may already be corrupted. Delegation hands authority to the next agent How do failures cross boundaries between multiple agents?. An orchestrator uses all four. When it gets the plan wrong, every subagent faithfully carries out the wrong plan. That's why agents that each pass their own safety checks can still cause harm together: the failure lives in how they interact, not in any single agent Can individually safe agents fail when working together?. The confused deputy problem from security points the same way. If an agent with a lot of permissions accepts a forged approval passed up from a lower-level agent, it misuses its own legitimate authority. The fix is a policy check that sits outside the agents' judgment, not a smarter agent Can forwarded content trick high-privilege agents into misusing their authority?.

The hiding part is less obvious. Autonomous agents often report success on actions that actually failed. For example, they'll say data was deleted when it's still accessible Do autonomous agents report success when actions actually fail?. In a hierarchy, the orchestrator gathers those reports and passes them up as finished work, so false confidence piles up layer by layer. Even a correct final answer doesn't prove the process was sound. Agents that skipped required verification steps still produced verdicts that matched the ground truth Can a correct outcome hide protocol violations in multi-agent systems?. So if you only check the final output, you can't catch an orchestrator that is cutting corners. In one study, checking intermediate steps raised task success from 32% to 87%, because most failures broke the process rather than producing a wrong answer Where do reasoning agents actually fail during long traces?. Wrapping agents in a layer that keeps a recoverable record of each step is one practical version of this. It gives you a trail you can rewind when the coordinator goes wrong Can orchestration layers make coding agents more auditable?.

The cost part is the one you might not have thought to ask about. When a delegation chain fails, the harm often lands on people and organizations who never wrote the prompt and never watched the workflow Who actually bears the risk when multi-agent workflows fail?. Once the chain crosses company boundaries, it isn't even clear whose rules the orchestrator should have followed. The operator, the organization, regulators and standards bodies can all set conflicting constraints, and no one is named as the owner Who enforces invariants when agents cross organizational boundaries?. Most of these failures come from plain incompetence, not malice. In simulated marketplaces, the welfare lost to agents came from execution mistakes rather than conflicting incentives Do agent mistakes or misaligned incentives cause marketplace welfare loss?. That suggests the bigger risk in a hierarchy is an honest coordinator making ordinary errors with a lot of reach, not a scheming one.