chart/AgenticOrgChart.com
menu
Pattern 04 / multi-level

Hierarchical Agents: Multi-Level Reporting Topology (2026)

A manager agent routes work to team-lead agents, who route work to worker agents below them. Documented in LangGraph as the hierarchical-teams primitive.

Hierarchical agents topologyA manager agent at the top routes work to two team-lead agents. Each team-lead routes work to two worker agents below it. The pattern is documented in LangGraph's hierarchical-teams primitive.Managertop-levelTeam-lead ATeam-lead BWorker A1Worker A2Worker B1Worker B2multi-level reporting + delegation
Hierarchical agents topology. A manager at the top delegates to two team-leads. Each team-lead manages two workers. Each layer is itself a supervisor pattern; the hierarchy is supervisor patterns nested inside supervisor patterns.Pattern documented in: LangGraph hierarchical-teams tutorial langchain-ai.github.io/langgraph/tutorials/multi_agent/hierarchical_agent_teams. Accessed 30 April 2026.

When hierarchy adds value

When sub-teams have clearly separable scopes. A research team and a writing team operating on the same goal benefit from each having its own team-lead supervisor that handles its team-specific routing, while a top-level supervisor handles only the team-to-team handoff. The org chart compresses cleanly into three layers: top, team-lead, worker.

When the manager’s reasoning load would otherwise exceed a single context window. A flat supervisor pattern with twelve workers requires the supervisor to hold all twelve worker capabilities in its prompt. Decomposing into three team-leads of four workers each lets each team-lead carry the four-worker prompt; the top-level supervisor only carries the three team-lead descriptions.

LangGraph hierarchical teams

LangGraph documents hierarchical-teams as a composed pattern: each team is itself a graph (a supervisor with workers), and the top-level graph treats the team as a single node. State is partitioned: the team’s internal state is private; the top-level state holds only the inputs and outputs of each team. Reference is the LangGraph hierarchical-teams tutorial at langchain-ai.github.io/langgraph/tutorials/multi_agent/hierarchical_agent_teams. Access date: 30 April 2026.

The same shape can be implemented in CrewAI by defining a hierarchical process where each manager is itself a Crew, and in AutoGen by nesting GroupChats. The structural property (multi-level supervision with partitioned state) is what defines the pattern; the specific framework primitive is implementation detail.

Reporting and accountability

The structural choice maps onto operational accountability. If a worker errs, the team-lead is the first responder; if a team-lead consistently errs, the top-level supervisor is accountable. The escalation pattern is the same shape that established management hierarchies use, which is why the topology reads naturally as an org chart.

For audit purposes, the hierarchy gives you log granularity. The top-level supervisor logs every team-to-team handoff; each team-lead logs every worker dispatch. A reviewer can read the top-level log to see “research team called writing team” without having to expand the worker-level detail unless needed.

Common failure modes

Excessive delegation depth. A manager defers to a team-lead who defers to another team-lead, with the original goal getting more diluted at each layer. The remedy is to fix the depth at design time and not allow runtime sub-team creation.

Cost explosion. Each layer adds context tokens. A three-layer hierarchy with three handoffs per layer can spend several times the tokens of a flat supervisor pattern. The remedy is to measure cost-per-task and reject the hierarchy if it does not actually deliver a quality improvement.

Loss of context across layers. By the time the original goal reaches a leaf worker, important constraints can have been stripped. The remedy is a typed state schema that propagates the original goal to every layer alongside the per-layer sub-goal.

Reference example

LangGraph hierarchical-teams topologyLangGraph's hierarchical-teams primitive: a top-level supervisor delegates to two team-level supervisors, each of which manages two worker agents. Each layer is a subgraph in its own right.Top supervisorroutes to teamsResearch team leadsubgraphWriting team leadsubgraphSearchwebSearchdocsDrafterv1Polisherv2
LangGraph hierarchical-teams: research team plus writing team. The top supervisor routes between two team-lead supervisors. Each team-lead supervises two specialised workers (search variants on the research side, drafter and polisher on the writing side).Source: LangGraph hierarchical-teams tutorial langchain-ai.github.io/langgraph/tutorials/multi_agent/hierarchical_agent_teams. Accessed 30 April 2026.

Related on this site

For the process flow view of a hierarchical agent system, see agenticswimlanes.com.