Logiciel Solutions Contact Us
Success Stories Tech News Investors Contact Us
whitepaper
8 min read

Multi-Agent Orchestration Patterns for the Enterprise.

Agentic orchestration is the runtime layer that decides which model, agent or tool acts next, what state it sees and when a human must approve, and the pressure on it is almost always to add more agents.

Anthropic measured that a single agent uses about 4 times the tokens of a chat interaction and a multi-agent system about 15 times, while its orchestrator with parallel subagents beat a single agent by 90.2% on an internal research eval.

Anthropic, Cognition and OpenAI all advise starting with one well-tooled agent anyway. That evidence comes from vendor engineering reports on their own systems, a UC Berkeley study of seven frameworks and Gartner forecasts.

It is not a survey of enterprise deployments, and the pattern catalog and checklist built on it are design guidance rather than measured results.

In depth

One Agent Is The Default And A Second One Needs A Business Case.

01

Multi-agent systems cost about 15 times the tokens of a chat.

Anthropic's production data puts a single agent at roughly 4 times chat-level token use and a multi-agent system at roughly 15 times, and token usage alone explained 80% of performance variance on the BrowseComp evaluation. The same orchestrator-workers system beat a single agent by 90.2% on Anthropic's internal research eval and cut research time by up to 90% on complex queries by running 3 to 5 subagents in parallel. Those gains were measured on breadth-first research, where subtasks are independent and the information exceeds one context window. Anthropic itself says domains with many inter-agent dependencies, including most coding, fit poorly today. The multiplier is only worth paying where the value of each run can absorb it.

In shortThe multiplier is only worth paying where the value…
02

The companies building agents advise starting with one.

Anthropic's "Building effective agents" tells teams to find the simplest solution and add complexity only when it measurably helps, and it separates workflows, which run LLMs and tools through predefined code paths, from agents, which direct their own process. OpenAI's practical guide says to maximize a single agent first and split only when prompts carry too much conditional logic or tools overlap. Cognition warned in 2025 that parallel writers make conflicting choices, because actions carry implicit decisions. Its April 2026 follow-up reports that the setups working today keep one writer while other agents contribute review or advice, and its Devin Review agent catches an average of 2 bugs per pull request, about 58% of them severe. Extra agents have earned their place as reviewers sooner than as writers.

In shortExtra agents have earned their place as reviewers so…
03

Many multi-agent failures trace back to system design.

The MAST study from UC Berkeley analyzed seven popular multi-agent frameworks over more than 200 tasks and defined 14 failure modes, with expert annotators agreeing at a kappa of 0.88. It groups them into three categories: specification issues such as ignoring role specs or termination conditions, inter-agent misalignment such as conversation resets and withheld information, and task verification failures such as premature termination or missing checks. The authors conclude that many failures come from design choices, so fixes usually require architecture changes as well as prompt changes. Anthropic reports a design failure from its own build, where an early version spawned 50 subagents for simple queries. Better prompts alone rarely fix a topology problem.

In shortBetter prompts alone rarely fix a topology problem
04

Interest in agents is running ahead of operational maturity.

Gartner reported a 1,445% rise in client inquiries about multiagent systems from Q1 2024 to Q2 2025, and predicts up to 40% of enterprise applications will include task-specific agents by 2026, up from under 5% in 2025. It also predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of cost, unclear value or weak risk controls. The protocols are settling under neutral governance: Anthropic donated MCP to the Agentic AI Foundation, a Linux Foundation directed fund, on December 9, 2025, and by April 2026 more than 150 organizations supported A2A. Those are forecasts and adoption counts, not measurements of how deployed systems perform. Demand for agents is not evidence that a given design works.

In shortDemand for agents is not evidence that a given desig…
The detail

Three Design Review Questions That Decide Cost And Risk.

None of this requires picking a framework first. The whitepaper stays above the framework layer and treats each design as a set of choices a review board can see on paper. Three questions surface most of them, and each one is answered by a single named artifact rather than a slide.

Zone · 01

Which pattern fits the task

Ask which of the ten cataloged patterns the design uses and why a single agent with workflows was not enough. Five of the ten are workflow patterns from Anthropic's taxonomy that keep control in code, and they are cheaper, faster and easier to test than patterns where an LLM directs the work. A design that cannot name its pattern cannot estimate its token bill either. The artifact is a written justification with the chosen pattern and its expected token multiplier.

Zone · 02

Who writes and what survives

Ask which agent owns each record or file, and what happens to a run when a process crashes or a human approval takes three days. Cognition reports that working setups keep writes single-threaded, and Anthropic resumes failed agents from checkpoints rather than restarting them. A run that cannot resume repeats work it already did, including actions with side effects. The artifact is a state map naming one writer per record and the durable engine holding run state.

Zone · 03

Where MCP ends and A2A begins

Ask which connections are agent to tool and which are agent to agent. MCP connects an agent to tools and data, while A2A connects independent agents owned by different teams, vendors or clouds. Tool definitions are costly too: Anthropic showed one workflow drop from about 150,000 tokens to about 2,000 when the agent called MCP tools through code. The artifact is a protocol boundary diagram with the tool scopes granted to each agent identity.

By the numbers

The figures that make it a board-level conversation.

15x
tokens a multi-agent system uses compared with a chat interaction, against about 4x for a single agent
90.2%
margin by which an orchestrator with parallel subagents beat a single agent on Anthropic's internal research eval
Over 40%
of agentic AI projects Gartner predicts will be canceled by the end of 2027 over cost, unclear value or weak risk controls
Inside the report

What you'll take away.

01

Step 1 - Write down why one agent with workflows is not enough

Treat a second agent as an architectural decision. Record the workload signals behind it, such as parallel subtasks or information larger than one context window, alongside a cost estimate and an evaluation plan.

02

Step 2 - Price every design per successful outcome, not per call

Anthropic's data puts multi-agent systems at about 15 times chat-level tokens. Estimate the multiplier for the chosen pattern and fan-out, then check it against what one successful run is worth to the business.

03

Step 3 - Keep one writer per record and store run state durably

Make other agents read-only, wrap every LLM and tool call as a retryable, idempotent step, and version prompts, tools and workflows together so in-flight runs finish on the version they started with.

04

Step 4 - Build an eval set from real tasks and tag failures with MAST

Anthropic began with about 20 real queries. Trace every agent and tool call with OpenTelemetry GenAI conventions and sort failures into specification, inter-agent misalignment and task verification categories.

Questions

Frequently asked.

What is agentic orchestration?

Agentic orchestration is the coordination layer that decides which agent, model or tool acts next in an AI workflow. It covers task decomposition, routing, shared state, retries, human approvals and stop conditions.

It can be built in plain code, in a workflow engine such as Temporal or AWS Step Functions, or as a supervising LLM agent, and it is what turns single model calls into auditable business processes.

When should you not use a multi-agent system?

Do not use one when a single agent with good tools can finish the task, when subtasks share most of their context, or when several agents would write to the same artifact.

Anthropic, Cognition and OpenAI all advise starting with a single agent. Multi-agent designs fit parallel, high-value work that exceeds one context window, such as broad research, where Anthropic measured its largest gains.

How much more do multi-agent systems cost than a single agent?

About 15 times the tokens of a chat interaction, against about 4 times for a single agent, according to Anthropic's production data on its research system. The real multiplier depends on the pattern, the number of workers, iterations and the size of tool definitions.

Anthropic cut one workflow from about 150,000 tokens to about 2,000 by calling MCP tools through code instead of loading every definition.

What is the difference between A2A and MCP?

MCP connects an agent to tools, data and prompts, while A2A lets independent agents discover each other and exchange tasks, messages and artifacts. MCP works vertically and A2A works horizontally. Both now sit under Linux Foundation governance, and more than 150 organizations supported A2A by April 2026.

Most teams need MCP first and add A2A when agents owned by other teams or vendors must collaborate.

Why do multi-agent LLM systems fail?

Mostly because of system design. The MAST study from UC Berkeley examined seven frameworks over more than 200 tasks and defined 14 failure modes in three categories: specification issues, inter-agent misalignment and task verification.

Examples include agents ignoring their roles, losing conversation history, withholding information and ending tasks early. The authors found fixes usually need architecture changes as well as better prompts.

What does the whitepaper include?

A framework-neutral catalog of ten orchestration patterns, each with use cases, failure modes and cost and latency notes. It also covers state and memory, A2A versus MCP, durable execution options, observability and guardrails, a token cost estimating model, a pattern selection matrix and a 14-point production checklist.

The matrix and checklist are design guidance for reviews, not measured results.

How far does the evidence in this whitepaper reach?

As far as its named sources and no further. The cost and performance figures come from Anthropic's reports on its own research system, the failure taxonomy from one UC Berkeley study of seven frameworks, and the adoption figures from Gartner forecasts and Linux Foundation counts.

It is not a survey of enterprise deployments and does not benchmark any particular company's agents.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send Multi-Agent Orchestration Patterns for the Enterprise straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Bring the workflow you are thinking of splitting across agents.

Bring one workflow you are considering for a multi-agent design, along with the volume and value of a typical run. We will work through which pattern fits it, what it is likely to cost in tokens and where a human checkpoint belongs. SECTION 7 - FAQ - 5 to 8 questions

Book an architecture review