KC Tech Enthusiasts - Agentic AI Design Patterns with Claude Code
At the KC Tech Enthusiasts Meetup, I presented some recently emerging design patterns for developing agentic AI systems based on research from Anthropic
Intro
- KC Tech Book Club
- KC Tech Enthusiasts
- Cayden Sommer
- Jesse Saunders
- Evan Harmon
- Intros
- Name
- Background
- What you’d like to get out of today?
Today
- Quick Talk
- AI
- Agentic AI
- Claude Code
- Quick Demo of Claude Code
- Share strategies, learn from others, Q&A, dive deeper
- …?
- Profit
AI
- LLM plateau
- Agentic AI
- Similar problems
- Hallucinations
- High level reasoning and understanding falls short
- Adding compute and huge context windows don’t fix it
- Frequently need to start over with fresh context
- Back to traditional engineering
- Designing complete systems
- Agentic AI
- Managing memory
- Efficiently and automatically tracking progress
- Feedback loops with clearly defined success/failures
- E.g. only allowed to change when it passes a unit test
Design Pattern for Agentic AI
- A way to use agents in a repeatable way that requires designing and setting up harnesses that reliably manage the memory use of your AI
- State
- Context/facts
- What counts as success/fail
- How to approach problems
- How to track progress
- (Sounds similar to software architecture patterns)
- Context Engineering
Claude Code
- Anthropic Research
- Honing in on this design pattern
- Initializer agent and coding agent
- Initializer agent sets the stage with the user prompt and how it sets the memory on other agents so that every other run with coding agents has dynamic context
- Why Claude Code has tooling that makes it easier to use this pattern
- .claude folder
- claude.md
- agent.md
- rules.md
- tools
- mcp
Demo
- how I’ve been using it
- what’s working well
- mowing bidder
- mcp
- Claude project instructions
- not working well
- zoho