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