Large Language Models are powerful, but a single prompt is just a function call. What if we applied the same principles that made software engineering successful (modularity, composition, separation of concerns, single responsibility) to agent design?

This session explores how the open-source Agent Package Manager (microsoft/apm) enables developers to author small, focused agent behaviours (primitives) and compose them into sophisticated systems. We'll walk through the journey from a single instruction file to a fully autonomous second brain: a knowledge system that work along any session, recalls before work, captures after work, ingests sources, answers queries to humans and agents alike, and compounds knowledge over time.

You'll see how:

- Single Responsibility produces primitives that do one thing well
- Composition over inheritance lets small skills combine into complex workflows
- Dependency management (just like npm/pip) makes agent capabilities portable and shareable
- Interface segregation keeps agent context windows lean and focused
- No slides-only theory. We'll dissect a real, production second-brain implementation: its wiki schema, ingest pipelines, query handlers, cross-session messaging, and the knowledge loop that ties it all together.

- Target audience: Software engineers, architects, and technical leads exploring agentic AI patterns. Familiarity with LLMs helpful but not required — software design experience is the prerequisite.
- Preferred duration: 45 minutes (30 min talk + 15 min Q&A), adaptable to 60 minutes with live demo.
Technical requirements: Screen share, no live internet dependency (demo uses pre-recorded flows as fallback).
- Session format: Conference talk / deep-dive.
- Conferences: Suitable for software architecture, developer experience, AI/ML engineering, and DevOps tracks.
- Key takeaway: Attendees leave with a mental model for designing agent systems using familiar engineering principles, and a concrete open-source reference implementation they can fork.