The enterprise AI conversation is changing. <\/p>\n\n\n\n
A year ago, many IT leaders were asking whether they should deploy an AI agent. Today, a more consequential question is emerging. How should organisations coordinate multiple agents working across departments, applications and business processes? <\/p>\n\n\n\n
Interest in this model is accelerating. Gartner reported a 1,445 per cent increase in enquiries about multi-agent systems between the first quarter of 2024 and the second quarter of 2025.\u00b9 Gartner also predicts that 40 per cent of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5 per cent in 2025.\u00b2 <\/p>\n\n\n\n
The architecture supporting this shift is multi-agent orchestration. It is becoming an important consideration for CIOs, CTOs and enterprise architects seeking to move beyond isolated AI experiments. <\/p>\n\n\n\n
A single AI agent can be effective when given a clearly defined task. It can interpret a request, use an approved tool and return a result. <\/p>\n\n\n\n
Most enterprise workflows, however, are not isolated. <\/p>\n\n\n\n
They cross system boundaries, involve different areas of expertise and depend on decisions made throughout a broader process. A customer request might require an agent to retrieve account information, another to assess contractual obligations and a third to\u00a0initiate\u00a0an operational workflow.\u00a0<\/p>\n\n\n\n
A multi-agent architecture divides this work among specialised agents with defined responsibilities. An orchestration layer coordinates how tasks are assigned, how context is shared and when work is transferred between agents.\u00a0<\/p>\n\n\n\n
The result is less like a single digital assistant and more like a coordinated team. Each agent performs a defined role while the wider system manages the workflow.\u00a0<\/p>\n\n\n\n
This division of responsibility can improve modularity and make individual components easier to test, govern and replace. It also introduces additional complexity. Microsoft\u2019s current guidance recommends starting with a single agent and separating responsibilities only when modularity, security boundaries or specialist capabilities make a multi-agent design necessary. <\/p>\n\n\n\n
Several recurring patterns are shaping the way multi-agent systems are designed. <\/p>\n\n\n\n
One agent completes a task before passing its output to the next agent. This approach is suitable for structured, linear processes such as document verification, invoice\u00a0handling\u00a0and compliance reviews.\u00a0<\/p>\n\n\n\n
Its strength is predictability. Each stage can be recorded and evaluated independently. Its weakness is that an error or delay early in the sequence can affect every later stage.\u00a0<\/p>\n\n\n\n
Multiple agents work at the same time on separate parts of a problem. Their results are then combined by an orchestrator or aggregation agent. <\/p>\n\n\n\n
This pattern can support research, security analysis, market assessment and risk evaluation. It reduces total processing time when tasks do not depend on one another, although conflicting results still require a method of reconciliation.\u00a0<\/p>\n\n\n\n
A supervisor or manager agent breaks a larger objective into smaller tasks and assigns them to specialist agents. <\/p>\n\n\n\n
This approach can make complex workflows easier to manage, but it places significant responsibility on the supervisor. Poor planning or task allocation at the top can reduce the quality of the entire system. <\/p>\n\n\n\n
An initial\u00a0agent assesses a request and transfers it to the most appropriate specialist.\u00a0<\/p>\n\n\n\n
This model resembles a service desk or contact centre. It can be effective when requests fall into recognisable categories, but routing rules and escalation paths must be carefully designed. <\/p>\n\n\n\n
An agent or group of agents repeatedly reviews and improves an output until it meets defined acceptance criteria. <\/p>\n\n\n\n
This pattern is useful for drafting, software testing, data validation and quality assurance. Clear stopping conditions are essential to prevent unnecessary processing, rising costs or endless revision loops. <\/p>\n\n\n\n
These patterns are not mutually exclusive. A single enterprise workflow might use a routing agent at the entry point, a supervisor within a business domain and sequential or parallel processing inside individual tasks.\u00a0<\/p>\n\n\n\n
The real architectural skill lies in combining these approaches without creating unnecessary complexity. <\/p>\n\n\n\n
One of the most significant developments in multi-agent architecture has been the emergence of open communication standards. <\/p>\n\n\n\n
Google introduced the Agent2Agent protocol, commonly called A2A, in April 2025. The protocol was designed to allow agents created with different frameworks or by different vendors to discover capabilities, exchange\u00a0information\u00a0and delegate tasks.\u00a0<\/p>\n\n\n\n
Google transferred the A2A project to the Linux Foundation in June 2025, establishing vendor-neutral governance for the standard. By April 2026, the Linux Foundation reported that more than 150 organisations supported A2A, with integrations across major cloud platforms and early production\u00a0use.\u00b3\u00a0<\/p>\n\n\n\n
This matters because most large organisations do not operate within a single technology ecosystem. <\/p>\n\n\n\n
An enterprise may use Microsoft for productivity and workflow automation, Salesforce for customer relationship management, Google Cloud or Amazon Web Services for infrastructure, and custom-built agents for industry-specific functions.\u00a0<\/p>\n\n\n\n
Without an interoperability standard, every connection between those systems could require a bespoke integration. A2A provides a common contract through which compatible agents can communicate while keeping their internal implementations private. <\/p>\n\n\n\n
A2A does not\u00a0eliminate\u00a0integration,\u00a0security\u00a0or governance work. Organisations still need to manage identity, authentication, authorisation and data exposure. It can, however, reduce the amount of custom communication logic\u00a0required\u00a0when connecting compatible agent platforms.\u00a0<\/p>\n\n\n\n
Microsoft is moving in the same direction. Copilot Studio now supports multi-agent orchestration and connections to external agents using the A2A protocol. Microsoft has also made computer-using agents generally available, allowing approved agents to interact with websites and desktop applications when suitable application programming interfaces are unavailable. <\/p>\n\n\n\n
The direction across the major platforms is becoming clearer. Enterprise agents are increasingly being designed as parts of coordinated systems rather than disconnected point solutions.\u00a0<\/p>\n\n\n\n
Deploying a multi-agent environment is not simply a matter of creating more agents.\u00a0<\/p>\n\n\n\n
Each\u00a0additional\u00a0agent introduces dependencies, permissions, context requirements and possible failure points. The overall system must therefore be designed around several architectural priorities.\u00a0<\/p>\n\n\n\n
When one agent transfers work to another, the receiving agent needs enough information to continue the task accurately. <\/p>\n\n\n\n
Passing too little context can produce incomplete or inconsistent outcomes. Passing too much can increase cost, expose unnecessary information or overwhelm the receiving agent.\u00a0<\/p>\n\n\n\n
Architects need to define what information moves between agents, where shared state is stored, how long it is retained and which agents are permitted to access it. <\/p>\n\n\n\n
Checkpointing can also help a system resume from a known stage after an error rather than restarting the entire workflow. <\/p>\n\n\n\n
A clear control mechanism is needed to assign tasks, manage dependencies and determine what happens when an agent fails.\u00a0<\/p>\n\n\n\n
This does not always require a single, central orchestrator. Some environments may use hierarchical or distributed coordination. Regardless of the model, the organisation needs an identifiable control plane for policy enforcement, access management and auditability. <\/p>\n\n\n\n
Without one, agents can duplicate work, issue contradictory actions or operate\u00a0beyond their intended responsibilities.\u00a0<\/p>\n\n\n\n
Traditional application monitoring focuses on infrastructure health, availability, response times and errors.\u00a0<\/p>\n\n\n\n
Agent observability must examine a broader set of questions. <\/p>\n\n\n\n
Did the system select the\u00a0appropriate agent? Was the necessary context transferred? Which tools and data sources were used? Why was a particular action taken? Did the workflow remain within its cost and risk limits?\u00a0<\/p>\n\n\n\n
Useful observability should capture agent interactions, tool calls, decisions, exceptions, latency, token usage and cost. It should also make the overall workflow understandable to technical teams and, where required, auditors or business owners. <\/p>\n\n\n\n
Each agent should have its own defined identity and permissions. <\/p>\n\n\n\n
Using shared or overly broad credentials makes it difficult to determine which agent performed an action and increases the potential impact of a compromised\u00a0component.\u00a0<\/p>\n\n\n\n
Access should follow least-privilege principles. An agent should only be able to retrieve or modify the information needed for its assigned function. <\/p>\n\n\n\n
As agents gain greater autonomy, governance becomes part of the technical architecture rather than a separate compliance activity. <\/p>\n\n\n\n
A minimum governance model should include: <\/p>\n\n\n\n
Human involvement should be based on risk rather than added indiscriminately. <\/p>\n\n\n\n
Routine, reversible and low-risk actions may be suitable for higher levels of automation. Decisions affecting employment, financial commitments, regulated information, customer rights or critical systems generally require stronger controls and human oversight. <\/p>\n\n\n\n
The economic potential surrounding generative AI is substantial, although it should not be attributed exclusively to AI agents. <\/p>\n\n\n\n
McKinsey estimates that generative AI could create between US$2.6 trillion and US$4.4 trillion in annual economic value across 63 use cases.\u2074\u00a0Agents may help organisations capture part of that opportunity by coordinating tools,\u00a0models\u00a0and workflows, but the estimate applies to generative AI more broadly.\u00a0<\/p>\n\n\n\n
McKinsey\u2019s 2025 global survey also found that AI high performers were more than three times as likely as other organisations to intend to use AI for transformative business change. More than one-third of these high performers reported allocating over 20 per cent of their digital budgets to AI technologies.\u2075 <\/p>\n\n\n\n
The same survey\u00a0demonstrates\u00a0that scaled agent adoption remains relatively early. No more than 10 per cent of respondents reported scaling agents in any individual business function.\u2075\u00a0<\/p>\n\n\n\n
This distinction is important. <\/p>\n\n\n\n
Multi-agent architecture is progressing quickly, but most enterprises are still developing the operating models, controls and technical foundations needed to use it at scale. Organisations should avoid treating agent deployment as proof of business value. Success depends on workflow redesign, reliable data, measurable\u00a0outcomes\u00a0and\u00a0appropriate governance.\u00a0<\/p>\n\n\n\n
Multi-agent orchestration is moving from research and experimentation into enterprise platforms and production environments. <\/p>\n\n\n\n
That does not mean every process needs multiple agents. A single agent or conventional automation may remain the simplest and most reliable option for many tasks. <\/p>\n\n\n\n
Multi-agent architecture becomes valuable when a workflow genuinely requires separate expertise, security boundaries, systems or decision stages.\u00a0<\/p>\n\n\n\n
The organisations best positioned to benefit will be those that design around clear business outcomes and\u00a0establish\u00a0the right foundations early. These foundations include interoperability, identity, observability, state management, cost\u00a0controls\u00a0and governance.\u00a0<\/p>\n\n\n\n
For IT leaders mapping an AI architecture, evaluating platforms or moving beyond isolated pilots, now is the time to determine where multi-agent systems add genuine value and where they add avoidable complexity. <\/p>\n\n\n\n