HOME/TECHNIQUE/Agentic Orchestration/Multi-agent topologies

TECHNIQUE

Multi-agent topologies

Agentic Orchestration

5APPLICATIONS
8OBSERVED OPERATORS
01

State of Practice

CROSS-VALIDATED — 11 OPERATORS

Across the cited deployed teardowns, multi-agent topologies are mostly built as specialist-role systems: operators split tasks into subagents/nodes, add routing or sequencing, and surround the topology with tool access, tracing, and validation layers.

Observed Practices

Decompose the workflow into specialized agents, roles, or graph nodes instead of one monolithic agent.

11 of 11 operators with cited teardown evidence in this pool.
DropboxRipplingSlackSiemensDoorDashAmazonAtlassianIberdrolaUbercubicAthena Intelligence

Put a router, classifier, director, supervisor, or master agent in front of specialists to choose the next subagent or expert path.

6 of 11 operators with cited teardown evidence in this pool.
DropboxRipplingSlackAtlassianIberdrolaDoorDash

Chain agent work in stages so one step’s output becomes the next step’s input.

6 of 11 operators with cited teardown evidence in this pool.
AmazonSlackDoorDashIberdrolaSiemensAthena Intelligence

Use parallel or concurrent delegation where the topology benefits from multiple simultaneous checks or subtasks.

3 of 11 operators with cited teardown evidence in this pool.
AmazonAtlassianDoorDash

Add an explicit validation role or falsification step before trusting agent findings.

4 of 11 operators with cited teardown evidence in this pool.
SlackDoorDashAmazoncubic

Scope context by domain, specialist, or task so agents do not load every tool, rule, or record into one prompt.

5 of 11 operators with cited teardown evidence in this pool.
DoorDashRipplingAtlassiancubicDropbox

Expose enterprise systems, repositories, data stores, or APIs to agents through tool interfaces, MCP servers, gateways, or similar controlled access layers.

7 of 11 operators with cited teardown evidence in this pool.
SlackIberdrolaUberAthena IntelligenceAtlassianDropboxRippling

Instrument multi-agent workflows with tracing, run logs, span visibility, or production monitoring.

6 of 11 operators with cited teardown evidence in this pool.
RipplingSlackIberdrolaAthena IntelligenceUbercubic

Use different model/cost levels for different agent roles or stages.

2 of 11 operators with cited teardown evidence in this pool.
SlackDoorDash

Where Operators Converge

Every operator with cited topology evidence decomposes work into multiple specialized agents, roles, or nodes.

Where Operators Diverge

Control-plane topology differs by operator: some use a central supervisor/director, some use staged pipelines or graphs, some expose an agent mesh, and some use a bundle of narrow micro-agents.

APPROACH 01

Central supervisor/director/router selects specialized agents or experts.

DropboxRipplingSlackAtlassianIberdrola

APPROACH 02

Staged pipeline or graph where outputs from one agent step feed later steps.

AmazonDoorDashSiemensAthena Intelligence

APPROACH 03

Agent mesh/platform where agents call other agents, tools, and systems.

Uber

APPROACH 04

Specialized micro-agents with narrow scopes for one product workflow.

cubic

Validation differs: operators use different mechanisms to keep multi-agent outputs usable and trustworthy.

APPROACH 01

Critic, disprove-it, adversarial, or grounded-execution validation inside the agent workflow.

SlackDoorDashAmazon

APPROACH 02

Tracing, eval sets, offline testing, or production monitoring to iterate on prompts and topology behavior.

Athena IntelligenceRipplingAtlassiancubic

Tool and execution integration differs across deployments.

APPROACH 01

Central MCP server or gateway mediates tool/data access and policy enforcement.

SlackIberdrolaUber

APPROACH 02

Tools and subagents are exposed through tool loops, schemas, or shared tool interfaces.

AtlassianDropboxAthena Intelligence

APPROACH 03

Agents execute in sandboxed or isolated environments for code, data normalization, or security testing.

DoorDashRipplingAmazon

Watch Items

Reliability, noisy outputs, false positives, and hallucination risk remain explicit operator concerns.

Context size, token consumption, latency, and cost require scoping, pruning, caching, or staged model allocation.

Debugging and provenance across agent hops are production concerns; operators add traces, invocation visibility, registries, and audit trails.

Consequential actions still have explicit control boundaries such as human approval, routing to human staff, or gateway policy enforcement.

02

Implementation Menu

CURATED DEFAULTS
NameKindMaturity
LangGraph multi-agent graphslibraryestablished
CrewAIlibraryemerging
AutoGenlibraryemerging
03

Observed in Production

5 APPS