Real systems + reference architectures
Real production multi-agent systems partners have built: with the internal agent breakdown drawn from public sources. Every row cites its engineering-blog or press announcement.
| System | Org | Industry | Topology | Framework | Composition | src | |
|---|---|---|---|---|---|---|---|
| Ask David | J.P. Morgan | Financial services | hub-spoke | LangGraph | 6 agents · 5 tools | HITL | ↗ |
| Multi-Agent Research System | Anthropic | AI research / consumer | swarm-dispatch | Anthropic internal (Claude Agent SDK precursor) | 3 agents · 3 tools | ↗ | |
| Investment-Research Multi-Agent Assistant | Amazon Web Services (reference architecture) | Financial services | hub-spoke | Amazon Bedrock Agents (multi-agent collaboration) | 4 agents · 4 tools | ↗ | |
| Klarna OpenAI Support Assistant | Klarna | Fintech / consumer | recursive | OpenAI Assistants API (custom orchestration layer) | 2 agents · 4 tools | HITL | ↗ |
| Bloomberg AI Assistant | Bloomberg | Financial data / media | hub-spoke | Custom (Bloomberg internal) | 4 agents · 4 tools | ↗ | |
| Genie · Uber Support Copilot | Uber | Consumer / mobility | recursive | Custom (Uber Michelangelo + LangChain components) | 2 agents · 3 tools | HITL | ↗ |
| Airbnb Support Summarisation | Airbnb | Consumer / travel | sequential | Custom (Airbnb internal AI platform) | 3 agents · 2 tools | HITL | ↗ |
| AlphaEvolve | Google DeepMind | AI research | swarm | Custom (DeepMind) | 3 agents · 3 tools | ↗ | |
| Magentic-One | Microsoft Research | AI research / consumer | hub-spoke | AutoGen | 5 agents · 3 tools | ↗ | |
| Agentforce Service Agent | Salesforce | Enterprise CRM | recursive | Agentforce (Atlas Reasoning Engine) | 4 agents · 5 tools | HITL | ↗ |
Ask David
Investment-research assistant for portfolio managers, analysts and traders: answers natural-language queries by querying structured databases, unstructured research documents and proprietary financial models.
Multi-Agent Research System
Powers Claude's Research feature, parallel sub-agent search across the web with source citation, used inside claude.ai for the depth-research mode.
Investment-Research Multi-Agent Assistant
AWS-published reference architecture for a multi-agent investment-research assistant on Bedrock: the template many banks and asset managers now base their internal builds on.
Klarna OpenAI Support Assistant
Customer-service AI assistant that handled ~2.3M conversations in its first month, two-thirds of Klarna's chat volume, doing the work of ~700 full-time agents.
Bloomberg AI Assistant
In-terminal assistant for Bloomberg Professional users: natural-language queries over 40+ years of proprietary financial data, news and research.
Genie · Uber Support Copilot
Internal assistant used by Uber's customer-support agents: surfaces answers from internal knowledge base and past resolutions to shorten handle time.
Airbnb Support Summarisation
Summarises long customer-support case histories for the agent before they engage: reduces handle time and improves first-response quality.
AlphaEvolve
Evolutionary coding-agent system that discovers novel algorithms: has produced verified improvements to matrix-multiplication and data-centre scheduling algorithms.
Magentic-One
Open-source generalist multi-agent system for complex tasks: browse, code, file navigation. Reference architecture behind Copilot Actions.
Agentforce Service Agent
Reference internal deployment on the Salesforce sales + support stack: customer-facing autonomous service agent grounded in Data Cloud + Knowledge.
How real teams wire orchestrators, sub-agents, tools and human gates
These architectures show a clear build pattern: one supervising unit, a small number of narrow workers, and a gate before anything irreversible. Businesses are converging on it independently because it is the arrangement that stays debuggable as the fleet grows.
What does a real multi-agent system look like?
Almost always the same arrangement: one orchestrator holding the plan, a few narrow sub-agents doing the work, a fixed set of tools they are allowed to call, and a human approval step before anything irreversible. Teams arrive at it independently because it stays debuggable as the system grows.
Which companies run multi-agent systems in production?
Ones with public write-ups include Anthropic's research system, Klarna's support assistant, Uber's support copilot, Bloomberg's assistant, Airbnb's ticket summarisation, DeepMind's AlphaEvolve and Salesforce Agentforce. Most are support, research or internal tooling rather than anything customer-facing and fully autonomous.
What is human in the loop?
A point where the system stops and waits for a person to approve, correct or take over. It is normally placed just before anything irreversible: sending money, emailing a customer, deleting records. Everything before the gate can run unattended.
What is an orchestration framework?
The software that coordinates the agents: it runs the steps, passes state between them and handles retries. LangGraph, AutoGen, Amazon Bedrock Agents and the OpenAI Agents SDK are common choices, and several large teams have written their own instead.
Do multi-agent systems cost more to run?
Considerably. Every agent adds its own model calls, and an orchestrator that reasons about each step pays for that reasoning too. Anthropic measured its multi-agent research system using around 15 times the tokens of an ordinary chat (Anthropic).
When should I split one agent into several?
When sub-tasks can genuinely run at the same time, or when one job needs a different tool set or a different model. Splitting for tidiness alone adds handoffs where context gets lost. A single agent with good tools is cheaper and easier to debug for most work.