Agentic AI Architecture: A Practical Guide to Building Smarter AI Systems

Agentic AI architecture is a system design that organises multiple specialised AI agents with each handling a distinct piece of a larger task, so the combined result outperforms any single model working alone.

"Essentially think of these LLMs as agents, and each agent is doing one piece of this task," explains Don Woodlock, President of InterSystems, in his Code to Care series.

Unlike traditional AI, which responds to a single prompt with a single output, agentic AI systems can plan, use tools, retain memory, and make autonomous decisions across multiple steps.

Instead of asking one model to do everything, you assign intelligent agents to specialised roles. One drafts. Another critiques. A third revises. Some aren't even language models; they're tools like search engines, APIs, and calculators. An orchestrator coordinates them all toward a shared goal.

What made this practical? Large language models crossed a reliability threshold. Earlier models couldn't be trusted with autonomous subtasks; they'd drift off-topic or fabricate constraints. Modern AI agents follow complex instructions well enough that each can be assigned a specialised role within a larger system, perceive its environment, and take action without human intervention at every step.

Large language models generate output one token at a time, moving only forward. There's no revision built into a single model call.

Think about how you'd actually build a marketing plan. You'd start with an outline. Write a draft. Review it. Get feedback. Revise. Maybe do a dry run, then produce a final version. Multiple passes, each improving the last.

An LLM does all of that in a single forward pass. No outline. No critique. No revision. The output is whatever the model predicted word by word, with no chance to step back and reconsider.

Agentic architecture fills this gap. By chaining agents together with each responsible for a different stage, you give AI systems the ability to draft, evaluate, and refine complex workflows that a single model can't handle alone.

Chain-of-thought prompting research ( Wei et al., 2022; Kojima et al., 2022) demonstrated that prompting LLMs to work through intermediate reasoning steps significantly improved accuracy on complex tasks. Agentic architecture extends that principle from a single prompt technique to a full system design: instead of asking one model to pause and think, you build a team of agents that inherently operates in stages.

Start with what Don calls "compound AI", stringing multiple LLMs together instead of relying on one. The simplest version uses three models in sequence:

1. Draft Agent: generates initial output
2. Critique Agent: evaluates the draft against the original requirements
3. Final Agent: revises based on the critique

"This actually works. This will give you a much better result. You can try it yourself actually with ChatGPT." - Don Woodlock

The key shift: stop thinking of each LLM call as "the AI." Start thinking of each as an agent with a role, a drafter, a critic, a finisher. You can use the same model for all three or mix different models for variety and perspective. The compound result outperforms any single call.

Once you think in terms of agents, the next realisation: not every agent needs to be a language model.

"Some of these agents could be tools, like do a Google search, call an API to schedule an appointment, a calculator API to do some maths." - Don Woodlock

In Don's example, a "data request agent" identifies what statistics would strengthen a marketing plan. A separate agent runs web searches to find them. The search agent is an API call. But within the system, it fills the same role: receive a task, execute it, return a result.

This is what separates agentic systems from prompt chains. The architecture accommodates any type of processing (e.g., language models, search engines, databases, calculators, or external APIs) all coordinated toward a single goal.

The final leap: the workflow doesn't have to be hardcoded.

Instead of defining a fixed six-step pipeline, you give an orchestrator agent a goal, a set of available agents, and general instructions. It decides the sequence. It might loop back for additional research. It might skip a step that isn't needed. It adapts.

Think of the orchestrator as a detective. A detective doesn't follow a checklist, instead they pursue a goal (solve the case) and let each piece of evidence shape what they investigate next. If a lead goes cold, they loop back. If a witness opens a new angle, they follow it. The orchestrator works the same way: goal first, sequence second.

This is what makes agentic architecture different from traditional automation. The system isn't following a script. It's pursuing a goal.

Every agentic system, regardless of complexity, shares four core modules plus an orchestration layer that ties them together.

The perception module acts as the system's sensory layer, gathering real-time data from APIs, databases, sensor feeds, or user interfaces and processing that data into usable representations. This is where natural language processing parses text inputs, computer vision interprets images, and data connectors pull structured records from enterprise systems. The perception layer sets the ceiling for everything downstream. An agent's output is only as good as the data it can access.

In healthcare, organisations using unified data platforms like InterSystems HealthShare, aggregate records across multiple EHR vendors through standards like FHIR^® and HL7®, give their agents a far more complete data foundation than those pulling from fragmented silos.

The cognitive module interprets inputs from the perception layer in light of the agent's current goals. This is where machine learning models, chain-of-thought logic, and planning algorithms operate.

The reasoning capabilities of large language models have made deliberative agents practical. They can weigh options, set sub-goals, and adjust plans mid-task.

But reasoning is also where things go wrong. A model that misinterprets data or sets the wrong sub-goal cascades errors downstream.

Memory systems give agents context beyond the current prompt. Short-term memory tracks active session context and intermediate results. What has the draft agent produced so far? What data has already been retrieved?

Long-term memory stores historical outcomes, user preferences, and accumulated knowledge, allowing the system to improve over time. Agents that maintain a record of past interactions and environmental states can perform long-horizon reasoning, which connects today's task to lessons learned from previous ones.

Context management across agents is one of the harder engineering problems in multi-agent setups; each agent's knowledge needs to be scoped correctly so it has what it needs without being overwhelmed by irrelevant information.

The action module executes decisions: generating text, calling an API, updating a database record, or handing off to the next agent. In Don's framework, the Draft, Critique, and Final agents each perform different actions in service of the same goal.

Actions can also trigger downstream processes in external software systems like scheduling appointments, filing tickets, updating CRM records, or pushing data to analytics pipelines. The action layer is where agentic AI connects to the real world.

The orchestration layer coordinates the flow of data and control between all other modules, deciding which agent runs next and what information it receives.

Beneath this sits the feedback loop, which is the mechanism that allows the system to evaluate its own progress, identify logic errors, and adjust strategies in real-time based on environmental feedback. Agentic systems improve over time through reinforcement learning loops as they ingest more data and refine agent behaviour with each iteration.

Agentic systems can be analysed at two levels: how individual agents are designed, and how multiple agents are organised into a system.

The way a single agent processes information and makes decisions shapes what it can handle:

Pattern	How It Works	Best For
Reactive	Responds to immediate stimuli without drawing on memory or prediction A model-based reflex agent maps percepts directly to actions	Simple, fast tasks: alert routing, threshold-based triggers
Deliberative	Reasons, plans, and uses internal models of the world to decide Slower but far more capable	Complex problems requiring multi-step planning: research, analysis
Cognitive	Mimics human-like thinking; including reasoning, learning, and autonomous decision-making with both short-term and long-term memory.	Tasks requiring adaptation: personalised recommendations, iterative content creation

Most modern AI agents are deliberative or cognitive. The reasoning capabilities of large language models make these patterns practical in ways that weren't possible five years ago.

How agents are organised determines how they collaborate and who makes decisions:

Pattern	Structure	Best For	Trade off
Single-agent	One agent making centralised decisions with tools and a system prompt Single agent systems handle the full task alone	Well-defined tasks: support lookups, document summarization, code review	Simple to build and debug Limited by one agent's capabilities
Pipeline (Vertical)	A leader agent oversees subtasks, directing multiple specialised agents in sequence Hierarchical AI agents report results back up the chain	Sequential workflows: content creation, data processing, report generation	Predictable and testable Inflexible if requirements shift mid-task
Peer (Horizontal)	Agents work as equals in a decentralised system, there is no single leader Agents operate independently and negotiate outcomes	Creative tasks: brainstorming, red-team/blue-team evaluation, interdisciplinary analysis	Diverse perspectives Harder to converge on a single answer
Hybrid	Combines structured leadership with collaborative flexibility, this balances leadership roles between structured and creative demands Leadership shifts based on task phase	Complex projects mixing structured execution and open-ended exploration	Most capable Most complex to build and debug

Don's "Draft-Critique-Revise" example is a pipeline where the workflow follows a predetermined sequence. His orchestrator example is a hybrid in which the system adapts based on what each step produces.

The ideal agent architecture depends on the requirements of the application. Start with a single-agent architecture for well-defined tasks. Graduate to multi-agent architectures only when task complexity demands it for things like specialisation, parallel execution, or cross-domain coordination.

In multi-agent architectures, the challenge is getting individual agents to work together.

Multi-agent collaboration requires solving three challenges: communication, coordination, and resource sharing.

Agents communicate through structured message passing. One agent's output becomes another's input. Communication protocols define the format, routing, and priority of these messages.

In vertical systems, agents reporting to a leader agent follow a hub-and-spoke pattern. In horizontal systems, agents negotiate peer-to-peer.

Standards like Model Context Protocol (MCP) were invented to standardise how agents connect to tools and data sources, reducing the integration burden for agent development teams.

One of the key features of multi-agent systems over single-agent architectures is parallel processing. Different agents can handle separate subtasks simultaneously (e.g., one researches while another drafts while a third queries a database).

This enables AI agents to solve complex problems faster than a single model working sequentially. But parallel execution introduces synchronisation challenges: agents need to merge their results coherently without contradictions or duplication.

The main challenge in multi-agent architectures is coordination. Without robust mechanisms for synchronisation, agents share resources inefficiently, duplicate work, or produce conflicting outputs.

This is especially acute in dynamic environments where task requirements shift mid-workflow. Performance monitoring at the system level (i.e., tracking which agents are bottlenecked, which are idle, and where errors accumulate) is essential for maintaining AI performance as the system scales.

Multi-agent systems are well suited for domains that require collaboration across diverse skill sets like compliance review, market research, workflow optimisation, or cross-functional projects where analytical, creative, and operational expertise need to converge on a single deliverable.

Agentic AI is not always the right choice. Summarising a document, classifying a support ticket, or translating text works fine with a single LLM call. Adding agents to those tasks creates overhead without meaningful improvement.

Agentic architecture earns its complexity when:

• Tasks require multiple steps: research, analysis, and synthesis in sequence
• External data is essential: the system needs to query databases, APIs, or live sources mid-workflow
• Self-correction improves outcomes: a critique loop measurably raises output quality
• Coordination across systems is needed: agents must pull from and push to different platforms

If the task has a single input, a single output, and no intermediate decisions, a well-crafted prompt will outperform an agentic system at a fraction of the cost and latency.

Agentic AI systems excel at solving open-ended problems that require autonomous decision-making and complex multi-step workflow management. In enterprise environments, this translates to measurable impact.

Organisations report business process acceleration of 30% to 50% when agentic systems handle task automation for operational workflows like compliance checks, data reconciliation, and customer onboarding.

In practice, this means agents tailored to specific business functions. As Don describes, you might build "a model that checks against your organisational policies or your brand", turning compliance review or brand governance from a manual bottleneck into an automated step in an agent-driven workflow.

Agentic AI can automate knowledge-intensive tasks that traditional AI models can't handle like analysing unstructured data across multiple sources, making judgment calls, and adapting to exceptions.

These systems function around the clock and scale to handle sudden spikes in data or inquiries without increasing headcount. Offloading repetitive cognitive work frees employees to focus on high-value strategy and creative innovation.

In healthcare, agentic AI enhances both operational efficiency and clinical decision-making. Agents that analyse medical images, cross-reference patient histories, and learn from new cases can improve diagnostic accuracy over time.

The integration of agentic AI into existing enterprise systems strengthens competitive advantages across sectors.

Platforms like InterSystems IRIS Data Platform provide the interoperability backbone these agents need in healthcare and beyond, connecting systems through standards like FHIR and HL7 regardless of vendor.

Supply chain operations benefit similarly. Agents that sense demand shifts and adjust fulfilment in real time need to run analytics on live transactional data, not on yesterday's warehouse extract. InterSystems IRIS supports this natively through translytical processing, running both transactional and analytical workloads on the same data without ETL delays.

Implementing agentic AI systems requires careful selection of agent types, tools, and agentic AI frameworks that align with specific AI capabilities and business objectives. The ecosystem has matured rapidly.

Agent frameworks handle agent logic, coordination, and tool use:

Framework	Maintained By	Approach
LangGraph	LangChain	Graph-based workflows. Agents as nodes, transitions as edges. Maximum flexibility, steeper learning curve.
CrewAI	Open source (AI Fund)	Role-based "crews". Assign agents roles, goals, and backstories. Fastest path to multi-agent prototypes.
Microsoft Agent Framework	Microsoft	Unified framework (merging AutoGen + Semantic Kernel). Deep Azure integration. Enterprise-focused.
Google ADK	Google	Code-first toolkit with built-in tools, Model Context Protocol support, and visual debugging UI. Integrates with Vertex AI.

These frameworks handle agent logic and coordination. But agents also need access to enterprise data that's spread across databases, interoperability engines, and real-time analytics. Platforms like InterSystems IRIS provide the data infrastructure layer that agents query and act upon, enabling developers to build agentic AI capabilities into existing software systems.

Selecting a practical framework is essential for building maintainable, scalable systems. The right choice depends on your team's expertise, your cloud environment, and whether you need pre-built agent patterns or maximum customisation.

Start here: Pick one framework and build the draft-critique-revise pattern from Leap 1. Use a single LLM for all three agents. Get it working. Then expand.

Building agentic systems introduces tradeoffs that single-model approaches avoid. Organisations must navigate significant technical complexity to ensure successful execution.

• Latency compounds. Each agent adds a round-trip. A three-agent pipeline takes roughly 3x the latency of a single call. Design for acceptable response times before adding agents.
• Errors propagate. A flawed output from the Draft Agent feeds into the Critique Agent, which may not catch it. Build validation checks at each handoff so they are not just at the end.
• Costs scale. More agents means more API calls means higher costs. Measure cost-per-task and compare against the quality gain. Sometimes a well-engineered single prompt is cheaper and good enough.
• Debugging is harder. When a five-agent system produces a bad result, you need to trace which agent went wrong. Structured logging at every agent boundary is essential.
• Diminishing returns hit fast. Adding a fourth, fifth, sixth agent rarely helps as much as improving the first three.

Scalability is one of the biggest challenges in agentic AI architecture, particularly when transitioning from single-agent to multi-agent systems. Agentic AI systems require robust infrastructure to handle higher workloads and ensure seamless communication among agents. What works for a three-agent prototype may collapse at twenty agents. Message queues overflow, context windows fill up, and coordination overhead become the bottleneck.

Testing agentic AI requires both synthetic and real-world datasets to ensure robustness and reliability. Synthetic data lets you stress-test edge cases and failure modes in controlled conditions. Real-world data reveals the messy, unpredictable inputs the system will actually encounter. Test each agent in isolation first, then test the full system end-to-end. The interactions between agents surface problems that unit-level testing misses.

The process of designing an agentic AI architecture is iterative. Architectures should be periodically reassessed as workload characteristics change. What started as a simple pipeline may need dynamic orchestration as task complexity grows.

As Don puts it, the goal is "building workflows where one or many of the agents could be directing what happens next, instead of you pre-thinking this whole flow in code or in your application." Start with the simplest architecture that works. Let experience guide the evolution.

Agentic AI architecture transforms AI from a single-prompt tool into a coordinated team of specialized agents.

The core insight, as Don frames it, is straightforward: start by thinking of LLMs as agents, recognise that not all agents need to be LLMs, and let the system direct its own workflow instead of hardcoding every step. Build a draft-critique-revise loop first. Add tools. Add orchestration. Let the system's complexity grow with your needs, not ahead of them.