Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
Design · Architectures

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.

Production case studies · 10 · scan the index, then read the diagram below
SystemOrgIndustryTopologyFrameworkCompositionsrc
Ask DavidJ.P. MorganFinancial serviceshub-spokeLangGraph6 agents · 5 toolsHITL
Multi-Agent Research SystemAnthropicAI research / consumerswarm-dispatchAnthropic internal (Claude Agent SDK precursor)3 agents · 3 tools
Investment-Research Multi-Agent AssistantAmazon Web Services (reference architecture)Financial serviceshub-spokeAmazon Bedrock Agents (multi-agent collaboration)4 agents · 4 tools
Klarna OpenAI Support AssistantKlarnaFintech / consumerrecursiveOpenAI Assistants API (custom orchestration layer)2 agents · 4 toolsHITL
Bloomberg AI AssistantBloombergFinancial data / mediahub-spokeCustom (Bloomberg internal)4 agents · 4 tools
Genie · Uber Support CopilotUberConsumer / mobilityrecursiveCustom (Uber Michelangelo + LangChain components)2 agents · 3 toolsHITL
Airbnb Support SummarisationAirbnbConsumer / travelsequentialCustom (Airbnb internal AI platform)3 agents · 2 toolsHITL
AlphaEvolveGoogle DeepMindAI researchswarmCustom (DeepMind)3 agents · 3 tools
Magentic-OneMicrosoft ResearchAI research / consumerhub-spokeAutoGen5 agents · 3 tools
Agentforce Service AgentSalesforceEnterprise CRMrecursiveAgentforce (Atlas Reasoning Engine)4 agents · 5 toolsHITL
Detailed architectures
J.P. Morgan · Financial services

Ask David

HITLhub-spokeLangGraph

Investment-research assistant for portfolio managers, analysts and traders: answers natural-language queries by querying structured databases, unstructured research documents and proprietary financial models.

OrchestratorSub-agentsTools + GatesHumanSupervisor AgentStructured Data AgentUnstructured Data(RAG) AgentAnalytics AgentLLM-as-JudgeHuman-in-the-LoopReviewer
Internal agents · 6
Supervisor Agent · agent: Central orchestrator, receives the query, manages short/long-term conversation history, delegates to specialised sub-agents.
Structured Data Agent · sub: Translates natural language into SQL / API calls against J.P. Morgan's databases; summarises highly formatted data.
Unstructured Data (RAG) Agent · sub: Vectorises + retrieves from emails, meeting notes, reports; standard RAG pattern over internal knowledge base.
Analytics Agent · sub: Interfaces with proprietary financial models + Python execution to run portfolio analytics and generate visualisations.
LLM-as-Judge · gate: Evaluates output quality before it reaches the end-user; reflects on hallucination + citation completeness.
Human-in-the-Loop Reviewer · human: For high-stakes decisions requiring absolute precision, the system flags the response for human review.
Tools + data
SQL databasesInternal APIsVector store (unstructured docs)Python code executionProprietary financial models
LangChain blog + JP Morgan public reporting on the Ask David multi-agent architecture↗ T2 · confirmed
Anthropic · AI research / consumer

Multi-Agent Research System

swarm-dispatchAnthropic internal (Claude Agent SDK precursor)

Powers Claude's Research feature, parallel sub-agent search across the web with source citation, used inside claude.ai for the depth-research mode.

OrchestratorSub-agentsLead Researcher(Orchestrator)Search Sub-agents (N)Citations Sub-agent
Internal agents · 3
Lead Researcher (Orchestrator) · agent: Decomposes the user question into sub-questions; spawns and coordinates parallel sub-agents; synthesises the final response.
Search Sub-agents (N) · sub: Each explores a specific sub-question in parallel: issues web searches, reads pages, extracts findings. N = 3–10 depending on query breadth.
Citations Sub-agent · sub: Runs after synthesis, attaches specific URLs to each factual claim in the final report.
Tools + data
Web searchWeb browsingInternal memory / scratchpad
Anthropic Engineering blog, 'How we built our multi-agent research system' (2025)↗ T1 · confirmed
Amazon Web Services (reference architecture) · Financial services

Investment-Research Multi-Agent Assistant

hub-spokeAmazon Bedrock Agents (multi-agent collaboration)

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.

OrchestratorSub-agentsSupervisor AgentFSI Research AgentQuantitative AnalyticsAgentData Automation Agent
Internal agents · 4
Supervisor Agent · agent: Bedrock supervisor coordinates the sub-agents and merges their outputs.
FSI Research Agent · sub: Retrieves industry news + reports via Bedrock Knowledge Bases + web search.
Quantitative Analytics Agent · sub: Runs code (Python) against market data using Bedrock's code-interpreter action group.
Data Automation Agent · sub: Extracts structured data from earnings PDFs + filings via Amazon Bedrock Data Automation.
Tools + data
Bedrock Knowledge BasesBedrock Data AutomationBedrock code interpreterMarket data APIs
AWS Machine Learning Blog, Part 3 of the reference investment-research assistant series↗ T1 · confirmed
Klarna · Fintech / consumer

Klarna OpenAI Support Assistant

HITLrecursiveOpenAI Assistants API (custom orchestration layer)

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.

OrchestratorHumanSupport AssistantFallback Human Agent
Internal agents · 2
Support Assistant · agent: Handles the customer conversation end-to-end: refund policies, order status, payment issues, disputes across 35+ languages.
Fallback Human Agent · human: Escalation path when the assistant flags low confidence or the customer explicitly requests one.
Tools + data
Order-management APIsPayments APIsRefund systemKnowledge base retrieval
Klarna press release (Feb 2024)↗ T1 · confirmed
Bloomberg · Financial data / media

Bloomberg AI Assistant

hub-spokeCustom (Bloomberg internal)

In-terminal assistant for Bloomberg Professional users: natural-language queries over 40+ years of proprietary financial data, news and research.

OrchestratorSub-agentsTools + GatesRouter AgentStructured DataSpecialistNews + ResearchSpecialistBloombergGPT (orGPT-4)
Internal agents · 4
Router Agent · agent: Classifies the query intent and dispatches to the correct data-source specialist.
Structured Data Specialist · sub: Runs queries against Bloomberg's structured feeds: prices, fundamentals, holdings.
News + Research Specialist · sub: Retrieves from Bloomberg News + proprietary research using domain-tuned retrieval.
BloombergGPT (or GPT-4) · tool: Underlying reasoning model, Bloomberg trained a 50B-param domain model but also uses GPT-4 depending on task.
Tools + data
Bloomberg data feedsBloombergGPTVector search over researchTerminal function APIs
Bloomberg + arxiv 2303.17564 (BloombergGPT paper)↗ T1 · likely
Uber · Consumer / mobility

Genie · Uber Support Copilot

HITLrecursiveCustom (Uber Michelangelo + LangChain components)

Internal assistant used by Uber's customer-support agents: surfaces answers from internal knowledge base and past resolutions to shorten handle time.

OrchestratorSub-agentsCopilot AgentKnowledge RetrievalSub-agent
Internal agents · 2
Copilot Agent · agent: Assists the human agent, retrieves policy documents, suggests responses, drafts replies.
Knowledge Retrieval Sub-agent · sub: RAG over Uber's support knowledge base and past tickets.
Tools + data
Internal KBPast-ticket vector storePolicy engine
Uber Engineering blog, Genie internal AI on-call copilot↗ T1 · confirmed
Airbnb · Consumer / travel

Airbnb Support Summarisation

HITLsequentialCustom (Airbnb internal AI platform)

Summarises long customer-support case histories for the agent before they engage: reduces handle time and improves first-response quality.

OrchestratorSub-agentsHumanSummariser AgentIngestion StepSupport Agent (Human)
Internal agents · 3
Ingestion Step · op: Pulls the case history + all associated messages + prior interactions.
Summariser Agent · agent: LLM-generated summary with structured sections: issue, prior actions, sentiment.
Support Agent (Human) · human: Reads the summary before engaging; can drill into original transcripts if needed.
Tools + data
Case-history APILLM (summarisation prompt)
Airbnb Engineering, LLM-powered support platform↗ T1 · confirmed
Google DeepMind · AI research

AlphaEvolve

swarmCustom (DeepMind)

Evolutionary coding-agent system that discovers novel algorithms: has produced verified improvements to matrix-multiplication and data-centre scheduling algorithms.

OrchestratorSub-agentsTools + GatesSelector / EvolverProposer Sub-agents(N)Evaluator Sub-agent
Internal agents · 3
Proposer Sub-agents (N) · sub: Each generates candidate code mutations using a Gemini-class model.
Evaluator Sub-agent · gate: Runs the proposed code against a verifier / benchmark and scores it.
Selector / Evolver · agent: Selects surviving candidates; controls the evolutionary population; drives the search.
Tools + data
Code execution sandboxVerifier / benchmark harnessGemini model
DeepMind blog, AlphaEvolve introduction (2025)↗ T1 · confirmed
Microsoft Research · AI research / consumer

Magentic-One

hub-spokeAutoGen

Open-source generalist multi-agent system for complex tasks: browse, code, file navigation. Reference architecture behind Copilot Actions.

OrchestratorSub-agentsTools + GatesOrchestratorWebSurferFileSurferCoderComputerTerminal
Internal agents · 5
Orchestrator · agent: Plans + tracks progress + delegates to specialised agents; maintains a task ledger.
WebSurfer · sub: Chromium-based web browsing sub-agent.
FileSurfer · sub: Navigates and reads local files.
Coder · sub: Writes Python code to solve the sub-task.
ComputerTerminal · tool: Sandboxed shell for executing generated code.
Tools + data
Chromium browserPython execution sandboxLocal filesystem
Microsoft Research blog + arxiv 2411.04468↗ T1 · confirmed
Salesforce · Enterprise CRM

Agentforce Service Agent

HITLrecursiveAgentforce (Atlas Reasoning Engine)

Reference internal deployment on the Salesforce sales + support stack: customer-facing autonomous service agent grounded in Data Cloud + Knowledge.

OrchestratorSub-agentsTools + GatesHumanService AgentAtlas Reasoning EngineFlow / Apex ActionsHuman Handoff
Internal agents · 4
Service Agent · agent: Handles the customer conversation end-to-end: grounded in Data Cloud + Knowledge, follows Flow-defined guardrails.
Atlas Reasoning Engine · sub: Salesforce's proprietary planner, decides which action or Flow to invoke next.
Flow / Apex Actions · tool: Deterministic business logic, CRUD, business rules, external system calls.
Human Handoff · human: Cases the agent can't resolve escalate to a human with full context.
Tools + data
Data CloudKnowledge baseFlowApexPrompt Builder
Salesforce Agentforce architecture documentation↗ T1 · confirmed
Field notes

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.

From the corpus, curated by Brandon Chaplin
Common questions
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.

Next in the learning path