Q&A

Agentic Design Patterns: When, Why, and How to Use Them

• By Pure AI Editors
• 09/03/2026

AI agents promise something fundamentally different from conventional software automation: instead of simply following a predefined sequence of instructions, they can interpret changing inputs, reason about what to do next, select tools and take actions on behalf of users or other systems.

Microsoft's AI agent adoption guidance describes an agent as software that uses generative AI models to interpret inputs, reason through problems and determine appropriate actions. Instructions, retrieval, actions and memory help give agents the context and capabilities to deal with ambiguity and multi-step work that can be difficult to express entirely through fixed rules.

That flexibility, however, comes with an architectural price. Microsoft's AI agent orchestration guidance recommends using the lowest level of complexity that can reliably meet a workload's requirements. A direct model call may be enough for a simple task. A single agent with tools can introduce dynamic decision-making without the coordination burden of multiple agents. Multiagent orchestration can handle more complicated work, but it also brings additional latency, cost, coordination overhead and failure modes.

Those choices are at the center of the upcoming Agentic Design Patterns: When, Why, and How to Use Them session at Live! 360 Tech Con 2026, where Jocelynn Hartwig, Director, AI Solution Engineering at Microsoft, will offer an architecture-focused framework for deciding when agentic patterns are actually worth using.

Jocelynn works with public sector organizations on practical, secure and mission-aligned AI adoption, helping federal teams translate emerging capabilities into business and operational value. Her work includes Copilot, agentic workflows, workforce transformation and responsible AI adoption, with an emphasis on moving organizations from experimentation toward production-ready AI solutions.

That production transition is where many of the most important agentic design decisions emerge. An architect has to decide how much control to give a model, what tools it may invoke, what state it should retain, how it should recover when something fails and when a person should be brought back into the loop. Microsoft's AI agent shared responsibility guidance, for example, recommends constraining agent scope, setting planning and iteration limits, applying least privilege to tools, auditing actions and requiring human approval for high-impact, irreversible or sensitive operations.

Jocelynn's perspective starts even earlier in the design process: before deciding which agentic pattern to implement, determine whether the problem should be solved by an agent in the first place. If a rule, function or conventional workflow can produce the required result consistently, introducing an LLM into the decision path can add variability and operational complexity without providing enough additional value.

For cases that do warrant agents, the session will examine the design decisions behind reasoning, tool use, state, human involvement, failure recovery and task coordination, along with the real-world tradeoffs involving latency, cost, security, predictability, governance and user trust. The goal is to give architects a practical mental model they can use to build agentic systems purposefully rather than adding autonomy simply because the technology makes it possible.

We caught up with Jocelynn to learn what inspired the session, why "Should this be an agent at all?" is such an important architectural question, when deterministic workflows remain the better choice and what tends to surprise teams when agentic systems reach production.

PureAI: What Inspired You to Present on This Topic?
Jocelynn: The agentic landscape is evolving incredibly quickly. While some architectural patterns have started to emerge, the pace of change and urgency to implement have led to more exposure to the latest tools and frameworks than to the underlying architectural principles. I wanted to focus on those principles and help architects think more deliberately about when and how to use agentic patterns.

What Is the First Question an Architect Should Ask Before Deciding to Use an Agentic Design Pattern?
Should this be an agent at all? Before introducing autonomy and nondeterminism, architects should determine whether the problem actually requires reasoning, adaptation, or dynamic decision-making, or whether a simpler deterministic solution would work.

When Is a Conventional Workflow a Better Choice Than an Agent?
When the process or outcome needs to be deterministic. If the steps are known, the rules are stable, and the expected output can be explicitly defined, a conventional workflow is typically simpler, more predictable, and easier to operate.

Which Agentic Design Decision Has the Greatest Impact on System Predictability?
How much autonomy you give the system. Guardrails, model parameters such as temperature, and the number of agents involved all influence predictability, but ultimately the more decisions an agent is allowed to make independently, the more variability you introduce into the system.

How Should Architects Decide When an Agent Needs a Human Approval Step?
Human approval should be introduced when a person remains accountable for the outcome, when an action has meaningful or irreversible consequences, or when the decision requires judgment that should not be delegated entirely to the system.

Which Tradeoff Tends to Surprise Teams Most When They Move Agentic Systems Into Production?
The operational complexity that comes with orchestration. A prototype may work well as a relatively simple agent, but production systems need mechanisms for managing state, failures, retries, observability, guardrails, and coordination. Teams often underestimate how important that orchestration layer becomes.

What Is One Sign That an Agentic Architecture Has Become More Complex Than the Problem Requires?
When an LLM is being used to make a decision or produce an output that could have been handled reliably with deterministic logic. If a rule, function, or conventional workflow can solve the problem consistently, adding an agent often introduces complexity without adding meaningful value.

Note: Those wishing to attend the session can save money by registering early, according to the event's pricing page. "Save $400 by registering by the Super Early Bird deadline of Sept. 25," said the organizer of the event, which is presented by the parent company of PureAI.