From Standalone Agents to AI Agent Systems

Standalone agent

InputA request or trigger

AgentModel, planning and tools

OutputOne task handled

Simple input-output · Limited context · Isolated operation

AI agent system

OrchestrationDistributes tasks, tracks dependencies, resolves conflicts

Creative agent
Audience agent
Bidding agent

Data and memory
Tools and APIs
Applications

One or more specialized agents · Shared context and memory · Orchestrated, governed workflow

From agent to system. A standalone agent wraps a model with planning, memory and tools for a task. An AI system combines models, one or more agents, infrastructure, data and the applications they interact with.

An AI agent can be defined as a task-oriented wrapper around a model, adding layers for planning, acting and adapting. An AI system is the whole setup: the models, one or more agents, the operational infrastructure, the databases and the applications they interact with. Its architecture shapes the system’s flexibility and reliability.

How Agents Handle Planning and Memory

The main classification of agents depends on how they handle planning and memory, which defines how they perceive and respond to their environment.

Reactive Agents

These agents operate on simple input-output logic. They respond immediately to triggers without using memory of past events or engaging in complex planning.

Strength: Fast and simple to implement.
Limitation: Cannot recall context or adapt their strategy over sustained interactions.

Deliberative Agents

These agents build and maintain an internal model of their environment, which lets them plan their actions. They are context-aware and can pursue long-term goals.

Strength: Capable of context-aware, planned decision-making.
Limitation: Slower and more computationally intensive than reactive agents.

Hybrid Agents

Context-dependent

Hybrid designs combine predefined responses with planning capabilities. They may suit tasks with both predictable steps and variable inputs, but introduce additional design and evaluation work. Select them when the task requires that combination.

Other Architectural Patterns

Several other design patterns offer structured approaches to building agent systems for complex tasks.

Layered Architecture

This pattern organizes an agent’s functions into a hierarchy. Lower layers handle tasks such as data collection and preprocessing; upper layers manage decision-making. The separation of concerns simplifies development and maintenance.

Blackboard Architecture

This design suits problems that require expertise from several domains. Different AI components, or “knowledge sources”, monitor a shared data structure (the blackboard) and contribute when they can help, like a team of experts solving a puzzle together.

Subsumption Architecture

This model is built from layers of basic behaviors that can override each other: for example, a lower layer can interrupt a higher-layer goal to ensure safety. Each layer works independently, which makes the system reactive and robust, but layers can occasionally produce conflicting actions.

Single-Agent or Multi-Agent Systems

An important design decision is whether to use a single agent or a team of specialized agents. The choice shapes the system’s capabilities and its complexity.

Single-Agent Architecture

One AI agent, powered by an LLM, handles all the tasks.

Strengths: Simpler to develop, manage and control. No coordination between agents.
Weaknesses: Can struggle with complex tasks or with access to a large number of tools.
Suited to: Straightforward, repetitive tasks, such as a dedicated media planning tool.

Multi-Agent Architecture

Specialized AI agents work together, with an orchestrator managing the overall process.

Strengths: Can split complex work across specialized agents, run tasks in parallel and use smaller models for specific steps.
Weaknesses: More complex to coordinate, test and debug.
Suited to: Complex tasks with several interdependent parts, such as media trading with multiple planning aspects.

Choose the simplest architecture that meets the requirements. Add separate agents only when evaluation shows a benefit that justifies the added coordination and maintenance.

Challenges of Multi-Agent Collaboration

Multi-agent systems introduce challenges that need to be addressed at the architectural level.

Key Challenges

Coordination and communication bottlenecks: Without standardized communication protocols, agents can struggle to align their actions, which leads to fragmented strategies, delays and poor performance.
Emergent and unpredictable behavior: Interacting agents can produce system-level behaviors that were not explicitly programmed, such as unintended outcomes, infinite loops or instability.
Debugging complexity: Tracing the root cause of a failure across nested, often opaque interactions between agents can be difficult and time-consuming.
Governance issues: Distributed decision-making blurs accountability: when an error occurs, it can be hard to determine which agent, or which interaction, was responsible. Biases in individual agents can also be amplified as they propagate across the system.

Architectural Responses

Several design strategies help reduce these risks:

Tool-augmented reasoning: Agents call external APIs and functions, so they can work with current data and real systems instead of relying only on the model’s internal knowledge.
Agentic loops (e.g., ReAct): An iterative “reason, act, observe” cycle lets an agent perform a step, check the result and decide on the next step, instead of failing on a single attempt.
Multi-agent orchestration: A dedicated orchestrator distributes tasks to specialized agents, monitors dependencies and resolves conflicts, which addresses coordination bottlenecks.
Governance-aware architectures: Control and oversight are built in: role-based access control so agents cannot exceed their authority, sandboxing to contain their actions, and audit trails so that decisions can be traced and reviewed.