Skip to content
Technology transformation expertise firm · ARDIANE Group Innovate, secure, succeed

Artificial intelligence

Who decides the next step? AI agent orchestration, from fixed code to handing over

In brief

AI agent orchestration refers to the way an application decides which agents run, in what order and who chooses the next step: the code, an orchestrator agent or the agent holding the task. Management Square compares these orchestration modes, what each costs in model calls, testing and traceability, and the situations in which each is the right choice.

Page contents

In the usage of development frameworks, AI agent orchestration refers to the way an application decides which agents run, in what order, and how the next step is chosen. The documentation for OpenAI’s Agents SDK, consulted on 12 November 2026, sums it up with these three questions. Vocabulary varies from one source to another: CrewAI’s documentation speaks of processes, LangChain’s of patterns.

Orchestration modes differ according to who decides. In a sequence fixed by code, the team has written the order of the steps. With an orchestrator, a main agent breaks down the request and delegates to specialised agents. In a handoff, the agent holding the task relinquishes it to another. The first suits cases where the steps are known in advance, the second cases where the breakdown changes with each request and a single agent must answer for the result, the third cases where a specialist must take over the exchange itself.

What AI agent orchestration decides, and who decides it

The notion of a multi-agent system refers to an architecture; orchestration, to the rule of circulation within it. It arises as soon as a process includes several steps entrusted to a model, whether those steps are simple calls or fully fledged AI agents.

Mode Who chooses the next step What the mode costs Situation in which it fits
Sequence fixed by code The code written by the team Cases to be anticipated one by one; an unforeseen request fails or goes to a person Steps known in advance, each testable on its own
Orchestrator and specialised agents A main agent, which keeps control Extra model calls; delegation instructions to write; a wrong breakdown can propagate Variable breakdown, a single agent answers for the result
Handoff The agent holding the task Transmitted context to keep under control; no agent holding the overall view; more trajectories to test A specialist takes over the exchange itself

This table is our synthesis of the documentation cited below, consulted on the same day; the costs in it are qualitative. To illustrate, let us take the after-sales service of a machine tool manufacturer, which receives requests for spare parts, breakdown reports and warranty questions.

When code fixes the sequence

The team writes the sequence of steps; the model only acts within each one. Anthropic’s guide “Building effective agents” (19 December 2024) describes this chaining as a series of steps in which each call to the model processes the output of the previous one, with programmed checks possible between two steps. CrewAI’s documentation calls the execution of tasks in a predefined order a sequential process. OpenAI’s Agents SDK documentation places in the same family the classification of a request followed by routing, the loop in which one agent evaluates another’s work, and the parallel execution of independent tasks.

For a parts request, the steps do not vary: identify the machine by its serial number, find the reference in the catalogue, check the stock, write the reply. Each can be tested on its own, with an input and an expected output, and a run can be reread in the order in which the code wrote it.

The price is rigidity: every case must have been anticipated. Anthropic considers this chaining suited to tasks that break down cleanly into fixed subtasks, and notes that it trades latency for accuracy, each call receiving a simpler task.

An orchestrator delegates to specialised agents

A main agent receives the request, chooses which specialists to call on and what to pass to them, then composes the answer from their returns. OpenAI’s guide “A practical guide to building agents” calls this pattern manager and considers it suited to cases where a single agent must keep control of the flow and of the exchange with the user. Anthropic intends it for complex tasks whose subtasks cannot be anticipated.

For a breakdown report, the breakdown of the work depends on the machine and the symptoms. The orchestrator queries an agent responsible for technical documentation, another responsible for the intervention history, then writes the proposed diagnosis for the technician.

LangChain’s documentation compares its patterns on examples. On a simple request, delegation adds a model call, since each result goes back through the main agent, and the delegated agents start from scratch at each request: their context stays isolated, but the path is repeated. In return, independent subtasks run in parallel. Added to this is design work: a specialised agent only knows of the request what the orchestrator passes to it.

This single point of passage also serves as a control point. OpenAI’s documentation notes that it makes it possible to apply common guardrails in a single place.

A handoff entrusts what follows to another agent

Here, no agent keeps control. A reception agent qualifies the request and passes it to a specialist, who becomes the point of contact. OpenAI’s guide speaks of a decentralised pattern, in which agents operate as peers, and considers it suited to cases where there is no need for a single agent to retain central control or synthesis.

In the example, the reception agent recognises a warranty question and hands over to the agent that knows the contracts. This agent continues the exchange with the customer, without the first having to relay each reply.

The cost lies in what is transmitted. In OpenAI’s Agents SDK, the agent receiving the handoff sees the whole conversation history by default, unless an input filter decides otherwise. What circulates from one agent to another is therefore set at design time, for the volume of context as for confidentiality. LangChain’s documentation notes that the active agent remains active from one turn to the next, which saves calls when a request is repeated, but that the handoff deals one after the other with domains that a delegation would have consulted in parallel.

The handoff also requires a ceiling on the number of transfers, otherwise two agents can bounce a request back and forth, and a trace that transfers do not interrupt, the subject of LLMOps. Since each agent can receive content read by another, the set-up is reviewed from the angle of AI agent security.

Choosing a mode, then changing it on the basis of measurements

We suggest asking the questions in this order. Are the steps known in advance? If so, the code fixes them. If not, is a single agent with the right tools enough? OpenAI’s guide recommends first pushing the capabilities of a single agent, and cites, among the signs that justify splitting, instructions that have become too branched and tools so similar that the agent confuses them. If several agents are needed, the choice between orchestrator and handoff depends on who must answer for the final result and speak to the user.

The modes can be combined: a sequence fixed by code can entrust just one of its steps to an orchestrator, which keeps the rest testable step by step.

Whatever the mode, agents act through the interfaces of applications, whose upkeep falls under information system architecture. The degree of autonomy, permissions and human oversight are decided before the orchestration mode: our page on AI agents describes them.

Moving from one mode to another is decided on measurements: the share of requests that fall outside the anticipated cases, the number of model calls and the cost per request, the share of results reworked by a person. As long as these measurements do not move, the most predictable mode stays in place.

Insights

Further reading

All insights
Permissions, tools, content read: AI agent security as a CISO examines it

AI agent security consists of limiting what an agent can reach, whatever the model decides: permissions reduced to the task, a closed list of tools, justified reads, content read treated as untrusted, a trace of every call and a tested shutdown. Management Square reviews the questions a security officer asks before authorising an agent, based on named public sources.

Read the analysis
How Article 6 of the European AI regulation identifies a high-risk AI system

A high-risk AI system is, under Article 6 of Regulation (EU) 2024/1689, a system linked to an Annex I product subject to third-party conformity assessment, or a system referred to in Annex III, unless the paragraph 3 derogation applies. Management Square reads here this article and its annexes in the consolidated version of 27 July 2026, and points out what the text leaves to be qualified.

Read the analysis
An AI centre of excellence: what for, and until when?

An AI centre of excellence is a cross-functional team that pools what several artificial intelligence projects have no interest in redoing separately: scarce skills, a technical foundation, an evaluation method. Management Square examines here the questions an executive committee asks before creating one: its usefulness, its reporting line, what it centralises, the risk of a bottleneck and its lifespan.

Read the analysis

FAQ

Frequently asked questions

Another question? Ask us

What is the difference between AI agent orchestration and a multi-agent system?

The term multi-agent system names an architecture, one in which several AI agents share a task. AI agent orchestration is the rule that sets, within that architecture or within a simple chain of calls, what runs, in what order and who decides what comes next. An orchestration fixed by code can therefore apply to steps that are not autonomous agents.

Can several orchestration modes be combined in the same application?

Yes. The documentation for OpenAI's Agents SDK and for LangChain, consulted on 12 November 2026, both state that their patterns can be combined. A sequence fixed by code can entrust one of its steps to an orchestrator, and an agent that has received a handoff can call other agents as tools. Each combination adds trajectories to test and to trace.

Is a development framework needed to orchestrate AI agents?

It is not a technical necessity. Frameworks such as LangChain, CrewAI, AutoGen or OpenAI's Agents SDK provide building blocks for calling models, defining tools and chaining steps. Anthropic's guide "Building effective agents" notes that they make it easier to get started, but that their layers of abstraction can hide the prompts and responses exchanged with the model, which makes diagnosis harder. The choice depends on the team's ability to read and trace what is running.

A project to bring to completion?

A first conversation is enough to understand your challenge, identify the right people and tell you frankly what we can bring.