SECURITY AI

Transparent and traceable

Reproducibility

TECNOLOGY

Artificial intelligence

AI-driven innovation to digitalize business processes and optimize productivity.

ERP Integrations

Comply platform’s AI Agents access company systems in accordance with the permissions granted.

RESOURCES

ESEMPI APPLICATIVI

Transport documents

Enhance data and optimize the management of transport documents.

Supplier Invoices

Automate supplier invoice management, from acquisition to recording.

Customer order management

Automate the acquisition, validation, and recording of customer orders in the ERP system.

Supplier order confirmations

Manage supplier order confirmations with AI agents.

FE INTERNAZIONALE

Italy

Poland

Slovakia

Spain

RISORSE

Compliance Calendar

The interactive calendar, always up to date for B2B compliance.

INNOVATIVE TECHNOLOGY

AI agentica

Find out how orchestrators coordinate AI agents and complex processes.

AI agents

Explore the technology behind our agents and how they work together toward a common goal.

Get in touch

Try and buy

We guarantee that we will have a solution tailored to your business up and running within three days.

COMPLANY

Certifications

Contact us

Newsletter

Let’s stay in touch

Sign up now for our free newsletter to stay up to date.

Agentic AI architecture: the multi‑agent ecosystem for modern enterprises

Simone Tiso

Simone Tiso

Chief Technology Officer

icona normalizzata

10/11/2026

icona conforme

(Follow-up to "The Anatomy of an AI Agent")

1. Why companies are shifting from "one agent" to "many agents"

The first generation of enterprise AI solutions relied on ever-more complex prompt engineering of LLMs to execute tasks. Early prototypes of agents were connected to tools and APIs via these prompts. While effective for prototypes or simple workflows, this model breaks down when exposed to real operational complexity: high volumes, exceptions, compliance rules, ERP integrations, and auditability requirements.

However, in practice, complexity is not solved by making a single model more powerful, but by distributing work across multiple specialized agents coordinated by an orchestrator.

  • Single-agent systems suffer from tool-selection errors, reliability issues, and degraded performance as complexity grows.
  • Decomposing tasks into subtasks improves accuracy, traceability, and overall output quality.

For business leaders, this translates into greater control, reliability, and scalability.

2. The architecture: an ecosystem of specialized agents supervised by an Orchestrator agent

2.1 The role of the Orchestrator agent

The Orchestrator is the supervisory agent responsible for:

  • interpreting the business objective(s) in the prompt;
  • breaking it into subtasks;
  • assigning tasks to the most suitable agents;
  • evaluating outputs;
  • determining follow-ups, retries, or escalation.

This model of Orchestrator-Worker pattern is currently widely recognized as the most robust for enterprise-grade systems.

2.2 Specialized agents

All agents have certain intrinsic functions, such as:

  • OCR → extract data from PDFs or images;
  • Search → retrieve information from databases or the web.

Beyond these functions, each agent is designed for a specific function. For instance:

  • ERP Agent → interact with SAP or other management systems;
  • Validation Agent → enforce business rules and compliance;
  • Synthesis Agent → generate structured outputs or summaries.

In an enterprise environment, the use of specialized agents with a reduced scope, and hence more focused on specific tasks, reduces hallucinations, improves tool usage, and increases reliability in production environments.

3. Multi-agent architectures: linear, cascading, shared-context, hybrid

Modern multi-agent systems can be configured in different ways depending on the company's requirements relating to agent autonomy, task complexity, and the company's governance requirements.

3.1 Linear agents

Linear pipelines are particularly effective when the business process is stable, highly structured, and has minimal branching logic. For example, a linear chain might include: extract data → clean data → validate fields → write to ERP.

Because each agent has a narrow scope and a clear input/output contract, linear architectures minimize ambiguity and reduce the cognitive load on the Orchestrator.

Linear pipelines are the most efficient choice when tasks do not require backtracking or complex decision-making. They also provide strong reliability and observability in production environments. However, linear agents are less suitable for workflows with frequent exceptions or dynamic decision paths, where more adaptive architectures (cascading or shared-context) offer superior resilience.

3.2 Cascading agents

In a cascading architecture, agents operate in sequence:

  • Agent A produces an output;
  • Agent B processes it;
  • Agent C validates it;
  • the Orchestrator Agent supervises and intervenes when needed.

This model is ideal for structured, repeatable processes such as order-to-delivery workflows or document onboarding.

3.3 Shared-context agents

In a shared-context architecture, each agent has partial visibility of the global objective and can take autonomous initiative.

The IBM research paper SOFAI-LM: A Cognitive Architecture for Building Efficient and Reliable Reasoning Systems with LLMs describes multi-agent systems, known as SOFAI (Slow and Fast AI), where agents share memory, context, and goals, increasing resilience and redundancy (IBM Research, 2025).

This configuration is useful when:

  • the process is unstructured;
  • resilience is required (agents can replace each other);
  • many exceptions must be handled.

3.4 Hybrid architectures

Many enterprise platforms adopt hybrid models that combine:

  • a central orchestrator;
  • specialized agents;
  • shared-context subsystems for complex reasoning or reconciliation tasks.

These architectures deliver both speed and reliability in complex enterprise workflows.

Hybrid systems pair fast, generalist LLM agents with slower but more rigorous specialist agents, such as rule-based validators, symbolic reasoners, or domain-specific solvers, to create a coordinated ecosystem in which each agent contributes its unique strengths. This approach excels in situations that require strict logic, deterministic constraints or domain-specific rules to be enforced.

In hybrid models, the orchestrator agent supervises the interaction between fast and slow agents. The orchestrator continuously evaluates the output of the LLM agents, detects inconsistencies or low-confidence reasoning, and triggers targeted interventions from specialist agents only when necessary. This selective escalation preserves speed while ensuring correctness. The orchestrator may also provide iterative feedback to the LLM agents, based on examples, constraints, or prior failures, allowing them to refine their reasoning without requiring additional fine-tuning.

When the LLM and feedback loop alone are insufficient, the orchestrator activates the specialist reasoning agent, supplying it with the contextual information accumulated during the LLM's attempts. This targeted activation improves overall performance while avoiding unnecessary computational overhead.

Academic evaluations (arXiv:2508.17959, 2025) across domains with very different requirements, such as graph-reasoning tasks and highly localized code debugging, demonstrate that hybrid multi-agent architectures can deliver both accuracy and efficiency.

By coordinating fast generalist agents with slower specialist agents, enterprises gain a system that is adaptable, resilient, and capable of handling both structured and unstructured tasks within the same operational framework.

4. How a multi-agent ecosystem works in practice

To understand the value of a multi-agent architecture, it helps to examine how the system behaves in a real operational workflow. Below is an expanded example showing how orchestrators and specialized agents collaborate to automate a complex, high-volume business process: in this case, the creation of a Delivery Note (DN) in SAP from supplier documentation.

This example illustrates not only the sequence of actions, but also the decision-making logic, error-handling mechanisms, and quality controls that make agentic AI reliable in enterprise environments.

Practical example: Automating delivery notes (DNs) in a manufacturing company

Business objective: optimize the logistics department by supporting users by accelerating data entry, improving data entry accuracy, and improving the management of delivery notes.

Step-by-step multi-agent workflow

1. Orchestrator interprets the business goal
The orchestrator receives the user request or system trigger (e.g., "New supplier PDF uploaded"). It determines the required workflow:

  • extract data;
  • normalize and map fields;
  • validate business rules;
  • create DN in SAP;
  • generate audit logs.

It then assigns tasks to the appropriate agents.

2. OCR Agent extracts unstructured data (such as the data in a PDF)
The OCR agent converts the unstructured PDF into machine-readable fields:

  • supplier name;
  • item descriptions;
  • quantities;
  • units of measure;
  • delivery date;
  • order references.

An external Judge Agent evaluates the extracted data to evaluate the extraction confidence. If this is low, the Judge Agent flags uncertainty for the Orchestrator Agent, which may trigger a retry with a different extraction model, a fallback to a more powerful OCR engine, or a human-in-the-loop check.

3. Normalization Agent maps extracted data to internal standards
Real-world supplier documents rarely match internal ERP conventions. Therefore, this agent:

  • converts free-text descriptions into internal item codes;
  • standardizes units of measure;
  • aligns supplier references with internal purchase orders;
  • resolves ambiguities using historical data or vector search.

If multiple possible matches exist, the agent returns options to the orchestrator, which decides whether to choose the highest-confidence match, request additional validation, or escalate to a human reviewer.

4. Validation Agent enforces business rules and compliance
This agent checks:

  • quantity tolerances;
  • price or discount rules (if applicable);
  • mandatory fields;
  • supplier compliance requirements;
  • VAT and fiscal rules;
  • internal quality or logistics constraints.

If inconsistencies are found, the agent provides structured feedback, such as "Quantity exceeds PO tolerance," "Missing supplier VAT number," or "Item code mismatch detected." The orchestrator then decides whether to correct the issue automatically, request clarification from the user, or halt the workflow.

5. SAP Agent creates the DN in the ERP
Once validated, the SAP agent performs the operational action: logs into SAP (via API or RFC), creates the DN, attaches extracted data, updates order status, and retrieves the DN number for confirmation. If SAP returns an error (e.g. locked document, missing master data), the orchestrator triggers a fallback workflow.

6. Audit Agent generates a complete trace of the process
For compliance and quality assurance, this agent produces a step-by-step log of all agent actions, extracted fields and confidence scores, validation results, SAP transaction IDs, timestamps, and explanations of any corrections or escalations. This ensures full transparency, essential for ISO, SOX, and AI Act compliance.

7. Orchestrator verifies the final outcome and closes the task
The Orchestrator agent performs a final check: did the output meet the original objective? Were all validation steps passed? Is the SAP record correctly created? If yes, the workflow is completed. If not, the orchestrator triggers corrective actions or alerts the user.

Why this example matters

This scenario demonstrates several key strengths of agentic AI:

  • Specialization: each agent performs a narrow, expert task.
  • Supervision: the orchestrator ensures quality and consistency.
  • Resilience: errors trigger retries, fallbacks, or escalations.
  • Auditability: every step is logged and explainable.
  • Scalability: multiple documents can be processed in parallel.
  • Compliance: validation agents enforce rules before actions reach SAP.

This is the difference between a "smart assistant" and a production-grade automation system.

5. Why this architecture is more reliable

A multi-agent architecture is not simply a more complex version of a single-agent system; it is a fundamentally more robust, governable, and enterprise-ready approach. The reliability gains come from structural design choices that reduce non-determinism, enforce quality, and ensure that every step of the process is observable and controllable. Recent academic research on agentic systems consistently highlights reliability as the primary advantage of orchestrated multi-agent ecosystems.

5.1 Reduced non-determinism through structured reasoning

LLMs are inherently probabilistic: the same prompt can produce different outputs. When a single agent is responsible for an entire workflow, this variability compounds and becomes difficult to control.

In a multi-agent system, however:

  • each agent performs a narrow, well-defined task;
  • the Orchestrator evaluates outputs at every step;
  • validation agents enforce rules and constraints;
  • errors are caught early rather than propagating downstream.

This transforms a probabilistic system into a supervised reasoning pipeline, significantly reducing the risk of hallucinations or inconsistent decisions. Breaking tasks into smaller reasoning units improves accuracy and reduces tool-selection errors.

5.2 Stronger governance and auditability

Enterprise environments require transparency: every decision must be explainable, traceable, and compliant. Multi-agent architectures naturally support this because:

  • each agent's actions are logged independently;
  • the Orchestrator maintains a complete audit trail;
  • validation agents enforce business rules;
  • human-in-the-loop checkpoints can be inserted at critical stages.

Governance becomes exponentially easier when responsibilities are distributed across specialized agents rather than concentrated in a single opaque model. This is especially important for Italian companies operating under strict fiscal, quality, and regulatory frameworks, and even more so with the EU AI Act.

5.3 Scalability through modularity

In a single-agent system, scaling means increasing the capacity of one model, which quickly becomes expensive and inefficient.

In a multi-agent architecture:

  • agents can be replicated independently;
  • workloads can be parallelized;
  • bottlenecks can be isolated and optimized;
  • new agents can be added without redesigning the entire system.

This modularity mirrors modern microservices architecture: each agent is a service with a clear contract.

As IBM's SOFAI-LM research shows, combining fast generalist agents with slower specialist agents yields better performance than relying on a single reasoning engine.

5.4 Robustness through redundancy and fallback strategies

A single-agent system has a single point of failure. A multi-agent system does not.

If an agent fails, the orchestrator can:

  • retry the task;
  • route the task to a different agent;
  • request clarification from the user;
  • escalate to a more powerful reasoning agent;
  • activate a fallback workflow.

This is particularly important in processes with high variability, such as document processing, logistics, or customer operations, where exceptions are the norm, not the anomaly.

5.5 Continuous quality improvement through feedback loops

Modern multi-agent systems incorporate metacognitive loops, where the orchestrator or evaluator agents:

  • assess the quality of outputs;
  • provide corrective feedback;
  • trigger refinement cycles;
  • escalate to specialist agents when needed.

This mirrors the hybrid reasoning approach described in arXiv:2508.17959 and IBM's SOFAI-LM architecture, where fast agents handle the bulk of the work and slower agents intervene only when necessary.

The result is a system that learns from its own errors in real time, without requiring model retraining.

5.6 Alignment with enterprise risk management

Finally, multi-agent architectures align naturally with enterprise risk frameworks:

  • segregation of duties → different agents handle extraction, validation, and execution;
  • least privilege → each agent has access only to the tools it needs;
  • traceability → every action is logged;
  • explainability → each agent's reasoning is isolated and inspectable;
  • compliance → validation agents enforce rules before actions reach ERP or financial systems.

This makes agentic AI a strategic enabler for safe, compliant, and scalable automation.

6. How to implement this in an enterprise: guidance for CIOs, COOs, and digital leaders

Implementing a multi-agent architecture is not a matter of "installing an AI model." It requires designing an operational ecosystem where agents, orchestrators, business rules, and enterprise systems work together reliably and safely.

For Italian companies, which often have complex supply chains, legacy ERP systems, and must operate within strict regulatory requirements, successful deployment depends on four pillars: infrastructure, security, integration, and continuous monitoring.

6.1 Build the right infrastructure: observability, state management, and scalability

Moving from a prototype to a production-grade agentic system requires a robust technical foundation. Multi-agent systems generate distributed workflows, where each agent may call tools, APIs, or other agents. Without proper infrastructure, this quickly becomes unmanageable. A production-ready environment should include:

  • Observability and distributed tracing: every agent action, tool call, and decision must be logged and traceable. This is essential for debugging, compliance, and continuous improvement.
  • State management: agents must share context reliably. This can be implemented through shared memory stores, vector databases, or orchestrator-managed state machines.
  • Dynamic scaling: some agents (e.g. OCR or ERP connectors) may need to scale horizontally during peak loads, while reasoning agents may require GPU-based scaling.
  • Versioning and controlled rollouts: new agent behaviours or updated prompts should be deployed gradually, with rollback mechanisms.

This infrastructure transforms agentic AI from a "smart assistant" into a mission-critical operational layer.

6.2 Ensure security, compliance, and governance from day one

Enterprise AI must operate within strict boundaries. Multi-agent architectures make this easier because each agent can be isolated and governed independently. Key requirements include:

  • Agent isolation and least-privilege access: each agent should only access the tools and data it strictly needs. This reduces risk and aligns with cybersecurity best practices.
  • Tool-level security controls: every tool call, whether to an ERP such as SAP, a database, or a document repository, must be authenticated, authorized, and logged.
  • Auditability and explainability: validation agents and orchestrators should produce human-readable logs explaining why actions were taken. This is essential for ISO, SOX, and AI Act compliance.
  • Alignment with the EU AI Act: the Act requires transparency, risk classification, and human oversight. Multi-agent architectures naturally support these requirements through modularity and traceability.

Security is a fundamental design principle, not an afterthought.

6.3 Integrate deeply with ERP, legacy systems, and operational workflows

The value of agentic AI emerges only when agents can interact with real business systems. This means building reliable connectors to:

  • SAP and other ERPs for creating documents, updating orders, retrieving master data, etc.;
  • SQL and NoSQL databases for structured data retrieval and validation;
  • document management systems for ingesting PDFs, invoices, transport documents, and compliance documents;
  • internal APIs and microservices for orchestrating cross-department workflows.

Integration agents must be deterministic, fully auditable, resilient to API failures, and aligned with existing IT governance. This ensures that agentic AI becomes a trusted extension of the enterprise architecture, not a shadow IT tool.

6.4 Implement continuous monitoring, evaluation, and improvement

IBM's SOFAI-LM research highlights the importance of metacognition: systems that evaluate their own reasoning and improve iteratively. In enterprise deployments, this translates into:

  • Quality evaluation loops: evaluator agents score outputs, detect anomalies, and trigger retries or escalations.
  • Performance dashboards: tracking accuracy, latency, error types, and business KPIs (e.g. time saved per transport document).
  • Human-in-the-loop checkpoints: for high-risk actions, such as financial postings or regulatory submissions.
  • Feedback-driven refinement: logs and errors are used to improve prompts, workflows, and agent behaviors without retraining models.
  • Automated fallback strategies: when LLM-based agents fail, specialist agents or deterministic rules take over.

Continuous monitoring ensures that the system becomes more accurate, more stable, and more aligned with business goals over time.

6.5 Treat agentic AI as a long-term capability, not a one-off project

Finally, successful companies treat agentic AI as a strategic capability, not a single deployment. This means:

  • establishing internal AI governance;
  • training teams to work with orchestrators and agents;
  • building reusable agent libraries;
  • expanding use cases progressively;
  • measuring ROI continuously.

The companies that win with agentic AI are those that build organizational readiness, not just technical readiness.

7. Conclusion: Agentic AI as strategic infrastructure

Companies adopting multi-agent architectures achieve:

  • higher reliability;
  • error reduction;
  • operational scalability;
  • governance and auditability;
  • smooth integration with existing systems;
  • fast ROI, thanks to agent specialization.

Agentic AI is a coordinated system of agents working toward a business objective, with an Orchestrator agent ensuring quality, control, and coherence.

About the Author

Simone Tiso

Simone Tiso

An expert in driving innovation and digital transformation, with over 15 years of experience in designing and implementing mission-critical infrastructures, primarily through an AI-driven approach. Proven ability to lead and coordinate multidisciplinary and international teams, aligning technology roadmaps with business objectives to maximize customer value and corporate competitiveness.

Newsletter

Subscribe to our monthly newsletter to stay up to date on the latest developments in business process automation.