AI agent orchestration: how to make agents work together

In short: AI agent orchestration means getting several models and tools to work on the same process, with a coordinator that decides who does what, in which order, and when a person steps in. It pays off when one agent can't cope with the complexity, specialisation or permissions involved; otherwise a single, well-connected agent is cheaper and makes fewer mistakes.

Key takeaways

• Orchestration is an architecture, not a product: a coordinator, specialised agents, tools connected through APIs, and control rules.
• Microsoft and Anthropic both advise starting with the simplest setup and adding agents only when you need them.
• Multi-agent systems burn far more tokens: in the case Anthropic described, about 15 times as many as a chat.
• Any action involving money, customers or personal data should go through a human checkpoint.
• The typical risks are errors passed from one agent to the next, hidden instructions in incoming text and agents with more permissions than they need: you manage them with minimal permissions, checks between steps and complete logs.

A company run like an orchestra of AI systems is a nice metaphor, but an online store or a B2B business needs a process that works every day. According to Istat, Italy's national statistics office, 16.4% of Italian companies with ten or more employees used at least one AI technology in 2025, up from 8.2% in 2024, rising to 53.1% among large companies. The next step is connecting the chatbots and assistants already in use to each other and to your core systems.

What is AI agent orchestration?

AI agent orchestration is the coordination of several agents and tools so that together they complete a process, with a control point that hands out the work and assembles the results.

An AI agent is a system in which a language model decides which steps to take and which tools to call to reach a goal, within limits you set. In its guide Building effective agents, Anthropic contrasts it with a workflow, where models and tools follow paths predefined in code.

The orchestrator receives the task, breaks it down, assigns it and decides when it is done: a model or a rule-based flow in n8n or Make. Tools are the functions an agent can use, such as reading an order, looking up a customer in the CRM or opening a ticket. The Model Context Protocol (MCP) is an open standard for connecting AI applications to external data and tools; the official documentation compares it to a USB-C port.

It is the next stage of business process automation: beyond "if A, then B" rules, some steps go to models that read, classify and propose decisions.

When do you need multiple agents, and when is one enough?

You need several agents when one can't cope with the instructions, the number of tools or the need to separate permissions; otherwise, one is enough. That is the guidance in Microsoft's guide to agent orchestration patterns: use the lowest level of complexity that reliably meets your requirements, because each level adds coordination overhead, latency and cost.

The numbers point the same way. Anthropic has described its multi-agent research system: on an internal evaluation it outperformed a single agent by 90.2%, but an agent uses roughly 4 times the tokens of a chat and a multi-agent system roughly 15 times. It works poorly when every agent needs the same context. In the same write-up, Anthropic lists where multi-agent setups shine: tasks with heavy parallel work, information that exceeds a single model's context window, and many complex tools to use. It adds that, as of June 2025, agents were not yet good at coordinating and delegating work to each other in real time.

Three questions help you decide:

• Does the process have genuinely different stages? Reading an email, checking the ERP and writing a quote need different tools.
• Do you need security boundaries? The agent answering customers shouldn't be able to issue refunds.
• Does the value justify the cost? If each case is worth a few euros, one agent or a rule-based flow will do.

What are the benefits of orchestration for a business?

The main benefit is splitting complex work into parts that are easier to build, check and improve one at a time. Microsoft's guide lists four advantages over a single agent that does everything:

• Specialisation: each agent focuses on one area, with shorter instructions and fewer misreadings.
• Scalability: you can add or change agents without redesigning the whole system.
• Maintainability: testing and fixes happen on one agent, not on a single monolithic block.
• Optimisation: each agent can use the model, data and tools best suited to its job, which also helps keep costs down.

Then there is speed on work that can be split up: in Anthropic's system, running 3 to 5 subagents in parallel, each using several tools in parallel, cut research time by up to 90% on complex queries. That figure comes from a research task, not a promise for every process.

For the business, the concrete gain is data moving between systems without copy and paste: if agents record the status of each case in a shared log, decision-makers see in one place information that used to be scattered across the store platform, the ERP and the CRM.

Where it applies in e-commerce and B2B

These are typical scenarios, to be assessed case by case:

• After-sales support: routing tickets, gathering order data and drafting the reply, as in the example below.
• Inventory and supply chain: one agent reads stock levels, one checks open purchase orders and one flags shipping delays; the reorder proposal goes to purchasing for approval. For the basics, see the guide to digitalising logistics in small businesses.
• Personalised marketing: one agent drafts message variants per customer segment and another checks them against brand rules and privacy consent before sending. In B2B the same pattern supports account-based marketing, with research on target accounts run in parallel.
• Sales and quotes: reading requests, pricing from the list and drafting a quote for the sales rep to sign off.
• Back office: matching orders, delivery notes and invoices, with anomalies flagged to a person.

For a wider view of how online selling between businesses is changing, see the future of B2B e-commerce.

What are the main orchestration patterns?

There are five basic patterns: sequential, parallel, routing, orchestrator-workers and maker-checker. Anthropic's guide covers them all; Microsoft's covers most of them, some under other names (concurrent, handoff, maker-checker).

Pattern	How it works	E-commerce example	B2B example
Sequential	Each agent works on the previous agent's output, in a fixed order	Product page: supplier data, copy, translation, SEO check	Quote: read the request, price it from the list, draft the document
Parallel	Several agents analyse the same input at once; results are merged	Reviews analysed simultaneously for quality, shipping and price	Target account researched across website, news and financials
Routing	A first agent classifies the request and hands it to the right specialist	Tickets split between returns, shipping and payments	Website enquiries split between sales, support and accounts
Orchestrator-workers	A central agent breaks the task down, delegates it and combines the answers	Weekly sales analysis with one sub-agent per channel	Meeting prep: company, stakeholders and CRM history
Maker-checker	One agent produces, another checks against explicit criteria and sends it back if needed	Product copy checked against banned claims	Tender responses checked against the stated requirements

Microsoft also describes two more advanced patterns. In group chat, several agents discuss in the same thread, guided by a manager that decides who speaks next; the guide suggests limiting it to three agents or fewer, because with more participants the flow gets hard to control. In the magentic pattern, a manager agent builds and updates a task list for open-ended problems with no predefined path; it is slow to converge and should be avoided when the work is time-sensitive.

In real projects they get combined: for many processes, routing up front, a sequential stretch and a final check are enough. Microsoft itself advises against forcing a workflow into a single pattern, for example a sequential first stage followed by parallel analysis.

How does it work on a real process?

Orchestration breaks each case into steps, gives each agent a narrow job and leaves the expensive decisions to a person. Take a hypothetical scenario: a homeware online store gets a few hundred after-sales requests a month by email and contact form, and one person handles them by opening the store platform, the warehouse system and the courier's site.

1. Intake. An n8n flow receives emails and form submissions and stores each request.
2. Routing. An agent running on a small, cheap model assigns a category (order status, return, damaged item, invoice) and an urgency level.
3. Data gathering. The specialist for that category reads the order, shipment status and returns policy. It has no access to payments.
4. Draft. A writer agent prepares the reply in the store's tone of voice, using only the real order data.
5. Check. A reviewer agent checks the rules: no refund promises, dates consistent with tracking, the customer's language. If something is off it sends it back once, then escalates to a person.
6. Approval. Informational replies go out automatically; returns and refunds reach the operator as drafts to approve. Everything is logged.

The same design works in B2B: for a components supplier, a quote request can be read by an agent that extracts part numbers and quantities, checked by a second agent against the price list and the customer's terms in the CRM, and turned into a draft the sales rep signs off. The agent doesn't decide discounts: it proposes them, showing the calculation.

Each agent needs a clear job, limited permissions and a rule for saying "done". To know whether it pays off, set your metrics up front, as covered in the guide to KPIs for measuring AI value.

How do agents talk to each other: MCP and A2A

Agents communicate in two directions: with tools and data, and with other agents. There is now an open protocol for each.

• MCP (Model Context Protocol) connects an AI application to external tools and data: databases, files, ERP, CRM. It is the "agent-to-tool" channel.
• A2A (Agent2Agent) connects different agents to each other. Google announced it in April 2025 with more than 50 technology partners, describing it as complementary to MCP. In June 2025 the project moved to the Linux Foundation, backed by more than 100 companies. With A2A, agents discover each other, exchange tasks and information and collaborate without sharing internal memory, tools or proprietary logic.

For a small or mid-sized business the practical question is different: you need A2A only if agents from different vendors have to talk, for example your CRM's agent and a custom-built one. If all agents live in the same n8n flow or the same framework, they pass data directly and the protocol isn't needed.

It also matters how much context passes from one agent to the next. Microsoft's guide notes that in multi-agent systems context grows quickly, because each agent adds its own reasoning and tool results: pass the next agent only what it needs, summarising or pruning the rest, to stay within model limits and keep costs down.

What tools do you need to orchestrate AI agents?

You need four building blocks: a workflow engine, language models, connectors and a database.

• Workflow engine: n8n or Make handle triggers, retries and approvals. In n8n, the AI Agent Tool node lets a main agent call other agents as tools.
• Models: a more capable one to orchestrate and write, smaller ones to classify and extract.
• Connectors: Shopify, WooCommerce, CRM and ERP APIs, or MCP servers. Without APIs, orchestration isn't reliable.
• State and logs: a database, Supabase for example, holding the status of each case and the human decisions.

For how these systems are reshaping software, see AI agents and the future of SaaS; more use cases are in the AI for business section. Designing flows and integrations is at the heart of my automation and AI projects for businesses.

Which frameworks and platforms are used to orchestrate agents?

The choice depends on who builds and maintains the system: visual platforms if the team doesn't write code, programming frameworks if there's a developer. The descriptions below come from each tool's official documentation.

Tool	Type	What it offers	When it makes sense
n8n	Visual automation platform	Flows with triggers, integrations and AI nodes; with AI Agent Tool, one agent calls others as tools	Small businesses and online stores that want to connect store, CRM and ERP without much code
LangGraph	Developer framework (Python)	Graph-based flows with shared state, fixed steps mixed with model-driven steps, durable execution, human-in-the-loop	Long-running processes that need precise control over state and flow
CrewAI	Developer framework (Python)	Crews of agents with roles, tools and memory, plus flows with persistent state	Technical teams that think in roles, such as researcher, analyst and reviewer
Microsoft Agent Framework	Developer framework (.NET and Python)	Successor to AutoGen and Semantic Kernel: multi-agent patterns, workflows with explicit paths, state management and human-in-the-loop steps	Companies already on Azure and the Microsoft ecosystem
OpenAI Agents SDK	Developer library (Python)	A few primitives: agents, handoffs between agents, input and output guardrails, built-in tracing	Projects already using OpenAI models that want a light start

None of these tools replaces process design: with the wrong pattern, even the best framework produces an expensive system. If you're interested in the open source side of automation, see also open source frameworks for automating work.

What are the risks of a multi-agent system, and how do you manage them?

There are four main risks: errors that spread, malicious instructions in incoming text, agents with too much power, and endless loops. Each can be reduced with specific design choices.

1. Compounding errors. Describing its research system, Anthropic writes that agents are stateful and errors compound: one wrong step can send the whole job off track. Microsoft recommends validating each agent's output before passing it on, and halting the flow when a response is incomplete or off-topic.
2. Hidden instructions in content (prompt injection). OWASP ranks it first among the risks of applications built on language models. In its indirect form, instructions inside a website or file the model reads change its behaviour; for an online store, the typical case is a customer email containing commands aimed at the agent.
3. Too much power (excessive agency). Again according to OWASP, the causes are unnecessary functions, excessive permissions on connected systems and too much autonomy; in multi-agent systems a harmful command can even come from another, compromised agent. An agent that only needs to read orders shouldn't be able to change or delete them.
4. Loops and runaway costs. Microsoft flags the risk of handoffs or review cycles that never end: you need a maximum number of attempts and a fallback, such as passing the case to a person.

The countermeasures Microsoft's guide recommends are the classic ones from distributed systems, applied to agents: timeouts and retries, least privilege for every agent and for the orchestrator, safety checks at several points (user input, tool calls, tool responses, final output), a log of every step, and tests for each agent and for the whole flow. On the protection side, see also AI and cybersecurity.

How to apply it in your company: first steps

1. Pick one process that is repetitive and measurable; note cases per month and minutes per case.
2. Map the steps: data, systems and who decides. Where there's a fixed rule, classic automation is enough.
3. Build a set of real test cases: Anthropic says it started with about 20 queries representing real usage, enough to see the effect of changes in the early stages.
4. Start with a single agent and add a second only where the first struggles with too many instructions or permissions.
5. Add human checkpoints for refunds, discounts, personal data and messages to new customers; cap automatic retries. Microsoft suggests deciding, for each checkpoint, whether human input is mandatory or optional and whether it is an approval or feedback sent back to the agent.
6. Measure for 30 days: time per case, drafts approved unedited, errors and model costs. Then decide. Since model output varies from run to run, judge quality with written criteria or a judge model, not word-for-word comparison.

Mistakes to avoid

• Starting from the architecture, not the process: five agents where one would do just add cost.
• Adding agents with no real specialisation: Microsoft lists this among common mistakes, together with the latency that grows with every hop between agents.
• Giving every agent the same permissions: the reply writer shouldn't be able to cancel orders.
• Using AI where a rule belongs, or a rule where AI belongs: a VAT calculation should be a fixed step; reading a free-form email is a job for a model.
• Letting several agents write to the same data in parallel: without a clear update order, the data becomes inconsistent.
• Ignoring data quality: with inconsistent customer records and price lists, agents spread errors faster, as covered in enterprise AI and operational context.
• Not logging decisions: without logs you can't tell which agent got it wrong or prove human oversight; see AI governance and explainability.

Frequently asked questions (FAQ)

What is AI agent orchestration?

It is the coordination of several AI agents and tools on the same process, led by an orchestrator that splits the work, assigns it and combines the results. The orchestrator can be a language model or a rule-based flow, for example in n8n or Make.

What is the difference between an AI agent and traditional automation?

Traditional automation runs fixed rules, such as "when an order arrives, create the invoice". An AI agent uses a language model to choose steps and tools, and it can handle unstructured input like emails and documents. The two usually work best together.

How many agents do you need to automate a business process?

Usually fewer than you'd think. Microsoft and Anthropic recommend starting with the simplest setup, often a single agent with its tools, and adding more only for stages that need genuinely different instructions, tools or permissions.

How much does a multi-agent system cost?

The cost items are design, integrations, model usage, workflow hosting and supervision. Model usage grows with the number of agents: in the research system Anthropic described, the multi-agent setup uses about 15 times the tokens of a chat. Small models for simple tasks keep this in check.

Can AI agents work with Shopify, WooCommerce and my CRM?

Yes, as long as those systems expose APIs or connectors and each agent has permissions limited to its task. If an MCP server exists for that system, the connection is simpler to build; without APIs, orchestration on that system isn't reliable.

Do agents' decisions need human oversight?

Yes, for refunds, discounts, outgoing messages and personal data. The EU AI Act requires human oversight for high-risk systems and transparency towards people talking to a chatbot; in other cases, too, approval checkpoints prevent the most expensive mistakes.

How long does it take to see results?

There's no timeline that fits everyone: it depends on the process, data quality and available integrations. Starting with a single process, after about 30 days of real use you already have the numbers to decide: time per case, drafts approved unedited, errors and model costs. Only then is it worth extending to other processes.

Sources

• Anthropic, “Building effective agents”
• Anthropic, “How we built our multi-agent research system”, June 2025
• Microsoft Learn, “AI agent orchestration patterns”, Azure Architecture Center
• Model Context Protocol, official documentation: “What is MCP?”
• n8n Docs, “AI Agent Tool” node
• Istat, “Imprese e ICT – Anno 2025” (Enterprises and ICT, 2025), press release
• European Commission, “AI Act: regulatory framework for AI”
• Google for Developers, “Announcing the Agent2Agent Protocol (A2A)”, April 2025
• Linux Foundation, “Linux Foundation Launches the Agent2Agent Protocol Project”, 23 June 2025
• OWASP Gen AI Security Project, “LLM01: Prompt Injection”
• OWASP Gen AI Security Project, “LLM06:2025 Excessive Agency”
• LangChain Docs, “LangGraph overview”
• CrewAI, official documentation
• Microsoft Learn, “Microsoft Agent Framework overview”
• OpenAI, “OpenAI Agents SDK” documentation

If you want to find out which process in your company is worth orchestrating first, and with how many agents, start with a free diagnosis.