HOME/TECHNIQUE/Agentic Orchestration/Multi-agent topologies

TECHNIQUE

Multi-agent topologies

Agentic Orchestration

4APPLICATIONS
7OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 11 OPERATORS

Observed deployments use multi-agent topologies as routed, role-specialized workflows: supervisors/directors/classifiers/graphs/trees/meshes delegate to subagents, with explicit context, tooling, and verification controls.

Observed Practices

Operators compose work across multiple agents, nodes, or agent-to-agent workflows instead of relying on one monolithic prompt.

11 of 11 operators in the teardown pool show multi-agent or multi-node orchestration evidence.
Alibaba CloudAmazonAthena IntelligenceAtlassianDoorDashDropboxRipplingRexeraSlackUbercubic

A routing or coordinating layer chooses which subagent, expert, branch, node, or downstream agent should handle the next step.

9 of 11 operators show an explicit coordinator, router, graph, branch, classifier, director, supervisor, or mesh coordination layer.
Alibaba CloudAmazonAtlassianDoorDashDropboxRipplingRexeraSlackUber

Operators specialize agents by role or domain: examples include red/blue teams, Director/Expert/Critic teams, read/RAG/action agents, Jira agents, reviewer/scout agents, and micro-agents for security, duplication, or editorial checks.

9 of 11 operators show role- or domain-specialized agents.
Alibaba CloudAmazonAtlassianDoorDashDropboxRipplingRexeraSlackcubic

Operators connect agents to tools or enterprise data systems, including MCP/tool-call interfaces, code/log/query tools, JQL execution, document retrieval, remote VMs, and downstream APIs.

11 of 11 operators show agents or subagents using tools, retrieval systems, data sources, or downstream APIs.
Alibaba CloudAmazonAthena IntelligenceAtlassianDoorDashDropboxRipplingRexeraSlackUbercubic

Several operators add a separate verification, critique, or falsification stage before trusting findings or making changes.

3 of 11 operators show a dedicated adversarial, critic, blue-team, or falsification step.
AmazonDoorDashSlack

Many operators instrument the topology with tracing, evaluations, production monitoring, offline eval sets, or acceptance metrics to debug multi-step behavior.

9 of 11 operators show tracing, monitoring, evaluations, or production/debug metrics around the agent topology.
Alibaba CloudAthena IntelligenceAtlassianDoorDashRipplingRexeraSlackUbercubic

Operators manage context pressure by pruning, scoped loading, external working memory, key-value observation stores, batching, or caching.

6 of 11 operators explicitly describe context-management mechanisms for multi-agent workflows.
Alibaba CloudAtlassianDoorDashRipplingSlackcubic

Where Operators Converge

Across the observed pool, the topology is not just multiple prompts: every operator connects agents or nodes to tools, data sources, downstream systems, or retrieval/action interfaces.

Where Operators Diverge

Operators differ in the orchestration shape they chose.

APPROACH 01

Supervisor, director, classifier, controller, or lead-scout routes work to subagents.

Alibaba CloudAtlassianDoorDashDropboxRipplingSlack

APPROACH 02

Graph, tree, DAG, or node workflow coordinates execution.

AmazonAthena IntelligenceAtlassianRexera

APPROACH 03

Agent mesh/platform supports agents calling other agents, tools, and systems.

Uber

Operators specialize agents around different units of work.

APPROACH 01

Security investigation or adversarial testing roles, such as Director/Experts/Critic or red-team/blue-team agents.

AmazonSlack

APPROACH 02

Enterprise domain or product-surface agents, such as read/RAG/action agents, Jira agents, domain experts, and subagents for complex work queries.

Alibaba CloudAtlassianDropboxRippling

APPROACH 03

Code-review pipelines split scouting, deep review, and narrow review checks across agents.

DoorDashcubic

APPROACH 04

Transaction-workflow agents or branches follow different parts of an operational process.

Rexera

Operators differ in how much autonomy they allow around writes or production changes.

APPROACH 01

Agents execute or prepare write operations inside product/workflow systems.

DoorDashRippling

APPROACH 02

Agents generate detection-rule improvements, but human oversight approves changes before production deployment.

Amazon

APPROACH 03

Agents focus on investigation, retrieval, reports, or quality checks rather than cited production writes.

Alibaba CloudAthena IntelligenceDropboxRexeraSlackcubic

Operators differ in how they keep multi-agent context bounded and inspectable.

APPROACH 01

Scoped loading or pruning reduces what the agent sees before routing or answering.

AtlassianDoorDashRippling

APPROACH 02

External memory, journals, artifact inspection, or key-value observation retrieval keeps long investigations from relying only on message history.

Alibaba CloudSlack

APPROACH 03

Caching manages overlapping context across specialized micro-agents.

cubic

Watch Items

Single-prompt or single-agent designs were reported as struggling with complex workflows, noisy comments, limited context, or too many tools.

Context windows and large intermediate outputs remain a recurring constraint in long-running or data-heavy agent workflows.

Operators report or mitigate hallucinations, false positives, false negatives, and agents taking the wrong path; several add critic, grounding, or falsification stages.

Multi-agent specialization can increase token or cost pressure, so some operators explicitly manage caching, model tiering, or cost metrics.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
LangGraph multi-agent graphslibraryestablished
CrewAIlibraryemerging
AutoGenlibraryemerging
03

Observed in Production

4 APPS