As interest in agentic AI grows, many teams jump too quickly from "LLM-powered feature" to "autonomous agent." The result is often unnecessary complexity, weaker control, and systems that are harder to test and maintain.
In practice, many successful AI applications begin with simpler patterns: a single LLM call, a routing step, a prompt chain, or a bounded workflow with clearly defined tool use. Agentic behavior can be powerful, but it should be introduced deliberately, where it adds clear value.
This talk presents a practical, software-engineering-focused approach to building production AI applications by starting with workflows before agents. We will look at how to choose the right level of autonomy for a problem, how to separate deterministic application logic from model-driven behavior, and how to apply orchestration patterns such as routing, chaining, and evaluation loops within a broader architecture.
We will also discuss how these patterns fit with structured outputs, guardrails, tool boundaries, and escalation paths, so that AI capabilities remain useful without becoming chaotic. The emphasis throughout is on architecture and implementation, not hype: how to build systems that are easier to reason about, safer to operate, and more maintainable over time.
Attendees will leave with a practical mental model for deciding when a workflow is enough, when an agent is justified, and how to build either one in a way that supports real production needs.

Key takeaways
- How to decide whether a feature needs a simple LLM call, a workflow, or a more autonomous agent
- Practical orchestration patterns for production AI applications
- How to keep control, safety, and maintainability in the surrounding software architecture

Target audience
Software engineers, architects, tech leads, and AI engineers building or planning AI-powered applications.

Suggested tags:
AI engineering, agentic AI, LLMs, software architecture, production AI, workflows

Compact Version:
As teams rush to add "agents" to their applications, many introduce more complexity than they actually need. In practice, the best AI applications often start with simpler patterns such as a single LLM call, routing, prompt chaining, or a bounded workflow with defined tool usage.
This talk presents a practical approach to building production-ready AI applications by starting with workflows before agents. We will examine how to choose the right level of autonomy, separate deterministic application logic from model-driven behavior, and design systems that are easier to reason about, test, and operate.
Attendees will leave with a practical framework for deciding when a workflow is enough, when an agent is justified, and how to build either one in a way that fits real software systems rather than just demos.