AI Agent Orchestration Platform Competition Concerns

1. Introduction

An AI agent orchestration platform is a system that coordinates multiple AI agents, models, tools, applications, APIs, data sources, and workflows to complete a user's objective.

Instead of:

User → One AI model → Answer

the architecture becomes:

User → Orchestration Platform → Multiple AI Agents → Tools/APIs/Data → Action

For example, an orchestration platform might decide:

which AI model should handle a task;

which agent should perform research;

which database should be accessed;

which external application should be called;

which payment or booking service should be used;

which agent's output should be trusted;

how different agents' outputs should be combined;

which final recommendation should be presented to the user.

This creates a potentially important new digital bottleneck.

The competition-law issue is not simply that one company operates an orchestration platform. The issue arises where a platform with substantial market power uses control over orchestration to exclude competing agents, favor its own services, restrict interoperability, exploit data advantages, or extend power into neighbouring markets.

Under Article 102 TFEU, dominance itself is not illegal; the legal concern is abusive conduct by a dominant undertaking. The European Commission's current Article 102 framework expressly focuses on exclusionary conduct and considers factors including market structure, entry barriers, buyer power and vertical integration. (Competition Policy)

2. What Is AI Agent Orchestration?

An orchestration platform acts as a manager of AI agents.

For example:

User request

"Plan my international business trip."

The orchestration system could automatically select:

Research Agent

↓

Travel Agent

↓

Translation Agent

↓

Hotel Agent

↓

Flight Agent

↓

Payment Agent

↓

Calendar Agent

The orchestration platform determines how these agents communicate and which services receive the user's request.

Therefore, the platform may control the route through which economic activity occurs.

3. Why Orchestration Can Create Competition Concerns

An orchestration platform may sit above numerous markets.

For example:

Foundation models

↓

Agent orchestration

↓

Search

↓

Travel

↓

Shopping

↓

Payments

↓

Cloud

↓

Enterprise software

If the same company controls several layers, it may potentially use power from one layer to advantage another.

This is particularly important because the Commission considers vertical integration among the factors relevant to assessing dominance. (Competition Policy)

4. The Orchestration Layer as a Bottleneck

Traditional software:

User → Application

Agentic software:

User → Orchestrator → Agent → Tool → Service

The orchestrator can therefore become the gateway.

It may determine:

"Which competitor gets the job?"

That decision can be economically significant.

Suppose five independent agents can provide the same service:

Agent A

Agent B

Agent C

Agent D

Agent E

The orchestration platform could technically send users to any of them.

If it consistently selects its own affiliated agent, the orchestration layer becomes a potential distribution bottleneck.

5. Main Competition Concerns

5.1 Self-preferencing

The platform may favor its own agents.

Example:

User: "Find the cheapest hotel."

Available:

Platform's hotel agent;

independent hotel agent;

independent travel agent.

If the orchestration system systematically sends the request to its own agent, the conduct may raise a self-preferencing issue.

The Google Shopping judgment is particularly relevant by analogy.

6. Case Law 1 — Google Shopping

Google and Alphabet v Commission

Case C-48/22 P
CJEU, 10 September 2024

The case concerned Google's general search service and its treatment of its own comparison-shopping service.

The Commission found that Google favored its own comparison-shopping results over competing services, and the CJEU dismissed Google's appeal against the General Court's judgment. (curia)

Principle

A dominant digital platform's method of presenting and ranking its own service can be relevant to Article 102 where the conduct has exclusionary effects.

Application to AI orchestration

Imagine:

Orchestration Platform

↓

User requests financial comparison

↓

Platform chooses its own financial agent

while competing agents receive fewer requests.

The legal analogy is:

Search ranking preference → Agent-selection preference

The technological mechanism is different, but the competition concern may be similar.

7. AI Agent Ranking as the New Search Ranking

Traditional platform:

"Which website appears first?"

AI orchestration platform:

"Which agent receives the task?"

This may become a major future competition issue.

Possible ranking criteria include:

price;

quality;

speed;

reliability;

affiliation;

commission;

advertising;

platform revenue;

user history.

If the platform secretly gives greater weight to its own commercial interests, competing agents may be disadvantaged.

8. 5.2 Interoperability Restrictions

An orchestration platform may control the technical interfaces through which agents communicate.

For example:

Agent A → Orchestrator → Agent B

If the orchestrator prevents Agent B from accessing necessary APIs, data or tools, competitors may struggle to participate.

This makes interoperability central.

9. Case Law 2 — Microsoft

Microsoft Corp. v Commission

Case T-201/04
General Court, 17 September 2007

Microsoft concerned, among other matters:

refusal to supply interoperability information;

interoperability between software systems;

tying Windows with Windows Media Player.

The General Court upheld important parts of the Commission's Article 82 decision. (Infocuria)

Application to AI orchestration

Suppose a dominant orchestrator controls:

agent APIs;

authentication;

communication protocols;

tool access;

context-sharing interfaces.

Independent agents may need those interfaces to compete effectively.

If the platform refuses interoperability under circumstances satisfying the applicable legal requirements, Microsoft provides a powerful analogy.

Example

Dominant orchestrator

→ controls API

→ denies rival agent access

→ rival agent cannot effectively participate

→ users remain within the dominant ecosystem.

10. 5.3 Refusal to Supply

An orchestration platform could control something essential such as:

access to users;

API access;

identity infrastructure;

proprietary agent protocol;

critical data;

tool marketplace.

A competitor might ask:

"Give my agent access to your orchestration layer."

A refusal does not automatically constitute abuse.

EU law applies demanding conditions to refusal-to-supply cases.

11. Case Law 3 — Bronner

Oscar Bronner GmbH & Co. KG v Mediaprint

Case C-7/97
CJEU, 26 November 1998

The case concerned access by a competing newspaper to a dominant undertaking's home-delivery network.

The Court rejected the abuse theory because alternative distribution methods existed and the relevant infrastructure was not shown to be indispensable in the required sense. (Infocuria)

AI relevance

Suppose a competing AI agent claims:

"We need access to the dominant orchestration platform."

The question would not simply be whether access is useful.

The analysis would ask whether:

the platform is genuinely indispensable;

viable alternatives exist;

access is realistically reproducible;

refusal would eliminate effective competition;

other legal conditions are satisfied.

Thus:

Useful platform ≠ legally indispensable platform.

12. 5.4 Critical Data and Agent Infrastructure

Some orchestration platforms may accumulate enormous datasets concerning:

user preferences;

successful agent workflows;

tool performance;

transaction outcomes;

agent reliability;

user feedback;

model performance.

The platform may then improve its orchestration decisions using those datasets.

This creates a potential:

Users → Interaction data → Better orchestration → More users → More data

feedback loop.

13. Case Law 4 — IMS Health

IMS Health GmbH & Co. OHG v NDC Health GmbH & Co. KG

Case C-418/01
CJEU, 29 April 2004

IMS Health concerned a specialized pharmaceutical sales-data structure and refusal to license access to it.

The Court considered when a resource can be regarded as indispensable for downstream competition. (Infocuria)

AI relevance

Imagine a dominant orchestrator possesses a proprietary:

agent-performance database;

tool compatibility database;

workflow architecture;

industry-specific knowledge graph.

A competitor may claim:

"Without this infrastructure, my agent cannot compete."

IMS Health demonstrates that the legal concept of indispensability is demanding.

The mere fact that an AI dataset is extremely valuable does not automatically make it an indispensable facility.

14. 5.5 Tying and Bundling

The orchestration platform might require:

"Use our model if you want to use our agent marketplace."

or:

"Use our cloud infrastructure if you want access to our orchestration API."

or:

"Agents using competing foundation models receive restricted functionality."

This could create potential tying or bundling issues.

15. 5.6 Cross-Market Leveraging

Suppose a company is powerful in:

AI orchestration

and also operates:

Cloud computing

Search

Advertising

Payments

Enterprise software

It could potentially use the orchestration layer to direct demand toward its other businesses.

Example:

User asks for accounting service

↓

Orchestrator

↓

Own accounting software

instead of

↓

Independent accounting software.

The concern would be whether this amounts to exclusionary leveraging rather than simply competition on the merits.

16. 5.7 Algorithmic Coordination

Agent orchestration can also raise Article 101 concerns.

Suppose several independent companies use the same orchestration infrastructure.

The system could potentially facilitate:

price coordination;

information exchange;

allocation of customers;

coordinated discounts;

common pricing rules.

This is not automatically unlawful merely because an algorithm is involved.

The legal issue remains whether the necessary elements of an agreement or concerted practice are established.

17. Case Law 5 — Eturas

Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba

Case C-74/14
CJEU, 21 January 2016

Travel agencies used a common computerized booking system. The system administrator imposed an automatic restriction on online discounts and sent a message concerning that restriction.

The Court considered whether the circumstances could establish a tacit agreement or concerted practice and addressed the evidentiary requirements. (Infocuria)

AI orchestration relevance

This is especially interesting for agentic systems because it demonstrates that computerized infrastructure can be part of a competition-law theory.

Imagine:

Several competing sellers

↓

Common AI orchestration platform

↓

Common algorithmic pricing rule

↓

Similar prices.

That similarity alone would not establish a cartel.

But if evidence shows that competitors knowingly participated in a common system or received and accepted information capable of coordinating their conduct, Eturas-type reasoning becomes relevant.

18. Case Law 6 — T-Mobile Netherlands

T-Mobile Netherlands and Others

Case C-8/08
CJEU, 4 June 2009

The case concerned information exchanged among competitors and the concept of a concerted practice.

AI relevance

Imagine competing businesses use an AI orchestration system that provides information about:

future prices;

customer demand;

inventory;

capacity;

strategic plans.

The orchestration system could become an information-exchange infrastructure.

Potential chain:

Competitors → Common AI system → Strategic information → Coordinated conduct

The critical issue remains whether the evidence satisfies Article 101 requirements.

19. 5.8 Agent Marketplace Foreclosure

Many orchestration platforms could develop an agent marketplace.

For example:

"Install agents from our marketplace."

The platform could control:

approval;

ranking;

certification;

visibility;

commissions;

API access;

user reviews;

distribution.

This creates a potential multi-sided market.

Side 1

Users.

Side 2

AI agents.

Side 3

Tool providers.

Side 4

Model providers.

The platform may therefore act as a gatekeeper between multiple groups.

20. Agent Certification as a Bottleneck

Suppose the platform says:

"Only certified agents may receive premium traffic."

Certification can have legitimate reasons:

safety;

cybersecurity;

quality;

privacy;

reliability.

But competition concerns could arise if certification is:

discriminatory;

opaque;

selectively applied;

excessively expensive;

designed to exclude rivals.

The key distinction is:

Legitimate quality control ≠ exclusionary certification.

21. Agent Commission Discrimination

Suppose:

The platform may argue that its own agent is cheaper to operate.

But if the difference is designed to make independent agents commercially unviable, the pricing structure could warrant examination.

Relevant analysis could include:

costs;

margins;

duration;

market coverage;

foreclosure;

efficiencies;

counterfactual pricing.

22. 5.9 Self-Preferencing Through Tool Selection

The platform may not explicitly rank its own agents.

Instead, it may control tool selection.

For example:

User asks:

"Find the best flight."

The orchestrator has access to:

Airline API A;

Airline API B;

Independent travel agent;

Platform's own travel service.

If the orchestration algorithm systematically sends requests to the affiliated service, the issue is not merely search ranking.

It is:

allocation of computational demand.

This could become one of the defining competition questions for agentic ecosystems.

23. 5.10 Data Advantage

Every agent interaction can produce information about:

which agent succeeded;

how long it took;

which tool worked;

which product converted;

which recommendation the user accepted.

The orchestrator may use this information to improve its own agents.

That produces:

Rival agents → performance data → platform learns → platform improves own agent → rival loses traffic.

This is a potential data feedback loop.

Whether it constitutes an infringement depends on the precise conduct and applicable law.

24. 5.11 Agent Switching Costs

Users may become dependent upon one orchestration platform because it stores:

agent preferences;

workflows;

credentials;

task history;

memory;

tool permissions;

personal context.

Switching to another platform may require rebuilding everything.

This creates:

Technical lock-in

The user's tools work only with Platform A.

Data lock-in

The user's history cannot easily be transferred.

Workflow lock-in

Custom workflows cannot easily be reproduced elsewhere.

Network lock-in

The user's preferred agents are available only on Platform A.

25. 5.12 Vertical Integration

An especially important scenario is:

Company A

owns:

foundation model;

AI agent;

orchestration platform;

cloud;

search;

marketplace.

This can create opportunities for:

tying;

self-preferencing;

foreclosure;

discriminatory access;

raising rivals' costs;

cross-subsidization.

The Commission expressly identifies vertical integration as a factor relevant to the assessment of dominance. (Competition Policy)

26. Agent Orchestration and Multi-Sided Markets

The platform can serve several groups simultaneously.

Users

Want:

reliable agents;

low price;

speed.

Agents

Want:

users;

visibility;

API access.

Developers

Want:

distribution;

tools;

data.

Model providers

Want:

computing demand;

customers.

Tool providers

Want:

transactions.

The orchestrator controls interactions between these groups.

This creates the possibility of cross-market leverage.

27. 5.13 Quality Degradation of Rivals

A sophisticated orchestration platform might not completely exclude competitors.

Instead, it could:

give competitors fewer tokens;

limit context;

provide slower APIs;

reduce tool access;

delay requests;

restrict memory;

place rivals lower in rankings.

This is sometimes more difficult to detect than outright exclusion.

Competition analysis may therefore require comparison between:

Actual treatment

and

Counterfactual neutral treatment.

28. 5.14 Interoperability and Agent-to-Agent Communication

Future AI systems may communicate directly:

Agent A ↔ Agent B

The orchestration layer could determine:

which agents can communicate;

what information they receive;

what permissions they have;

which APIs they can call.

This makes interoperability potentially more important than traditional website interoperability.

The Microsoft case provides an important conceptual precedent for considering interoperability as a competition issue in a dominant technological ecosystem. (Infocuria)

29. 5.15 Refusal to Interoperate

Suppose a dominant orchestration platform says:

"Our agents may communicate with our own applications but not with competing applications."

Possible questions:

Is the platform dominant?

Is access indispensable?

Are alternatives available?

Is competition being eliminated?

Is there an objective justification?

Is the restriction proportionate?

This is where Bronner and IMS Health become important analogies. (Infocuria)

30. 5.16 AI Agent Market Concentration

The market could potentially evolve into:

Many AI agents

↓

Few orchestration platforms

↓

One dominant orchestrator

This is significant because competition may shift from:

Which AI model is best?

to:

Which platform controls access to all AI agents?

The orchestration layer may therefore become more strategically important than any individual model.

31. 5.17 Raising Rivals' Costs

A dominant orchestrator could potentially increase competitors' costs through:

high API charges;

certification fees;

compute fees;

data-access fees;

transaction commissions;

integration costs.

An independent agent might technically remain available but become commercially unviable.

This can create a raising-rivals'-costs theory.

32. 5.18 Margin Squeeze

Consider:

Upstream: orchestration/API access

Downstream: AI agent service

The platform competes downstream with independent agents while charging them high upstream prices.

Example:

Platform's own agent

Cost of orchestration = €1

Independent agent

Orchestration fee = €8

The independent agent may be unable to compete.

This creates a potential margin-squeeze analysis analogous to established EU competition jurisprudence.

33. 5.19 Predatory Pricing

The platform could offer:

"Free orchestration forever."

Independent competitors may be unable to match the price.

But a free service is not automatically predatory.

Authorities would need to examine:

costs;

strategy;

market structure;

duration;

exclusionary effects;

recoupment where legally relevant;

efficiencies.

34. 5.20 Cross-Subsidization

The company could finance free orchestration through profits from:

cloud computing;

advertising;

payments;

enterprise software;

marketplace commissions.

Thus:

Profitable Market A

↓

funds

Free AI Orchestration

↓

weakens

Competitors in Market B

This may require detailed economic analysis rather than assuming that free pricing itself is unlawful.

35. 5.21 Consumer Choice

The ultimate competition question may involve:

fewer agents;

less innovation;

higher prices;

reduced quality;

lower privacy;

less choice.

But orchestration can also create substantial benefits:

lower search costs;

easier access to specialist agents;

improved productivity;

automatic comparison;

faster transactions.

Therefore:

Centralization is not automatically harmful.

The legal analysis must distinguish efficient coordination from exclusionary coordination.

36. 5.22 Evidence Required

An AI orchestration investigation could require unusually technical evidence.

Platform evidence

source code;

system architecture;

ranking rules;

agent-selection rules.

API evidence

access logs;

latency;

rate limits;

pricing.

Agent evidence

rejection rates;

ranking;

traffic;

commission.

User evidence

switching rates;

user dependence;

complaints.

Economic evidence

market shares;

entry barriers;

foreclosure;

pricing;

counterfactual outcomes.

Internal evidence

business plans;

emails;

strategy documents;

communications concerning rival agents.

37. Important Distinction: AI Output vs Competition Evidence

An AI system selecting one agent over another does not automatically prove anticompetitive intent.

The investigation must distinguish:

Legitimate optimization

Agent A is genuinely better, cheaper or safer.

from:

Potential exclusion

Agent A is selected because it belongs to the platform.

The relevant evidence could include:

algorithmic rules;

internal documents;

controlled experiments;

output patterns;

pricing;

traffic allocation;

counterfactual simulations.

38. Case Law Summary

These are analogical authorities, not cases specifically decided on AI agent orchestration.

39. Seven Major Competition Theories

1. Self-preferencing

Platform selects its own agents.

2. Refusal to interoperate

Competitors cannot connect.

3. Data foreclosure

Competitors cannot access important data.

4. Tying

Access to orchestration requires use of another service.

5. Raising rivals' costs

Independent agents pay substantially more.

6. Algorithmic coordination

Common orchestration facilitates competitor coordination.

7. Vertical leveraging

Power in orchestration is used to control neighbouring markets.

40. AI Orchestration Monopoly Formula

Orchestration Power

Users + Data + Agents + APIs + Network Effects + Switching Costs + Vertical Integration

↓

Platform Gatekeeper Position

↓

Control Over Agent Selection

↓

Self-Preferencing / Interoperability Restrictions / Tying / Discrimination / Foreclosure

↓

Potential Article 102 Concern

↓

Dominance + Abuse + Competitive Effects + Causation + Objective Justification

41. The Most Important Legal Distinction

A useful examination sentence is:

Control over an AI orchestration layer is not itself an abuse of dominance; the competition-law issue arises where a dominant platform uses that control in a manner capable of excluding competitors, restricting interoperability, leveraging market power into adjacent markets, or otherwise producing an unlawful exclusionary effect.

The Commission's Article 102 framework expressly separates dominance from abusive conduct and considers market definition, entry barriers, buyer power, resources and vertical integration when assessing dominance. (Competition Policy)

42. Six Cases to Remember for Exams

1. Google Shopping — C-48/22 P

Remember: AI agent ranking / self-preferencing.

2. Microsoft — T-201/04

Remember: Agent interoperability.

3. Bronner — C-7/97

Remember: Access is not automatically a legal entitlement.

4. IMS Health — C-418/01

Remember: Indispensable data/infrastructure.

5. Eturas — C-74/14

Remember: Computerized systems and algorithmic coordination.

6. T-Mobile — C-8/08

Remember: Information exchange between competitors.

Conclusion

AI agent orchestration platforms may become a new layer of digital market power because they can control not merely information but the allocation of tasks among competing AI agents and the routing of users toward commercial services.

The central transformation is:

Traditional platform:

"I decide what you see."

AI orchestration platform:

"I decide which agent acts for you."

That difference can be economically significant.

The most important future competition questions are therefore likely to concern agent selection, interoperability, API access, data advantages, self-preferencing, agent marketplaces, switching costs, vertical integration, algorithmic coordination and cross-market leveraging.

The existing jurisprudence already provides the main legal building blocks: Google Shopping for preferential digital treatment, Microsoft for interoperability and tying, Bronner/IMS Health for access to potentially indispensable infrastructure, and Eturas/T-Mobile for computerized coordination and information exchange. (curia)

One-line exam formula:

AI Agent Orchestration Competition Problem = Gatekeeper Control + Agent/Data/API Dependency + Network Effects + Switching Costs + Exclusionary Conduct → Potential Foreclosure and Article 101/102 Competition Concerns.