Spec First, Agents Second: Engineering AI Systems with Discipline

AI-assisted development has dramatically accelerated how we build software. Tools like Claude Code, MCP servers, Copilot, and local LLMs make it easy to generate code and spin up agentic workflows. But speed without structure leads to drift, fragile systems, and loss of architectural intent.

In real-world production environments, where performance, collaboration, and long-term maintainability matter, "vibe coding" does not scale.

This session introduces a practical, open source approach to AI-Assisted Engineering grounded in three ideas: specification-first development, product-minded engineering, and disciplined use of AI tools.

Using the open source project SpecKit, we’ll explore how structured, version-controlled specifications establish clear system intent before agents generate implementation. Specs become durable artifacts that guide contributors, constrain LLM behavior, and preserve architectural clarity.

We’ll demonstrate how:

* Specs shape and bound agentic workflows

* Product-minded engineering connects requirements to real user outcomes

* AI becomes a force multiplier when embedded in structured processes

Rather than replacing engineering discipline, AI works best when paired with explicit intent, traceability, and shared understanding. This talk is for developers and maintainers experimenting with agentic systems who want a repeatable, open-source-friendly model that balances speed with rigor.