Experiments in Agentic Coding

Name: Experiments in Agentic Coding
Start: 2026-10-02T15:15:00-05:00
End: 2026-10-02T16:00:00-05:00
Location: The Rand

When: Fri, Oct 2 · 3:15 – 4 PM
Where: The Rand, 3rd Floor

Coding agents can behave in surprising ways for both experienced and new devs. Devs new to agentic coding may have frustrating experiences trusting code that agents generate, or be using agents in sub-optimal way. This talk demonstrates how agents work, why they sometimes do funny things, and how to put guardrails on Claude. The demonstrations will use prompt engineering experiments with personas to solve a strawman problem that all devs know and love -- FizzBuzz. First, silly personas (e.g., pretend you're a goose) will be used to solve FizzBuzz, and demonstrate phenomena like non-determinism, constitutional AI, and context. Next, it will progress to a more practical prompts (e.g. Django novice, Java expert, Senior Architect), and a more realistic example, by adding new requirements to FizzBuzz. This second set of demos will show how intrinsic training biases or reward hijacking can lead your agent to generate bad code. For Python specifically, a plain agent, not configured properly, will default to generating messy and unreadable tests. Along the way, it will show practical tips for using Claude (plan vs. edit mode, context management, memories, skills and sub-agents), and demonstrate what kind of prompts are needed to generate clean code. Sample skills and sub-agent MDs will be provided for users to try with their own code.

Tech & Builders circuit, sponsored by

All of PySanAntonio II The rest of the day Everything at The Rand