Every founder has built an AI demo that worked beautifully once. Far fewer have built one they’d let run unattended on Monday morning.
That gap isn’t about model quality. It’s about structure, verification, and knowing which parts of your business an agent should touch at all.
This hands-on workshop takes you from prototype to production-shaped in a single hour. You’ll build a working agentic workflow end to end: it accepts messy real-world input, decides what to do with it, calls a tool, returns structured output your code can act on — and is checked by a lightweight evaluation harness before it goes anywhere near a customer. We use customer triage as the working example, but the pattern transfers directly to onboarding, research, document processing, and internal operations.
Along the way we tackle the decisions that determine whether an AI feature survives contact with reality: when to use an agent versus plain deterministic code, how to break a business process into agent-sized steps, how to design for non-determinism and silent failure, and how to know your system has degraded before your customers tell you.
You leave with running code, a reference architecture, and a delegation checklist you can apply to your own workflow the same week.
Key Takeaways
- Working Code: A complete, resilient agentic workflow running locally on your machine.
- Reference Architecture: Production-shaped structure covering agent, tool invocation, structured output, and evaluation.
- Delegation Framework: A clear framework for determining what belongs in an agent vs. what belongs in deterministic code.
- Reusable Eval Harness: A 20-line evaluation harness you can adapt and reuse the same day.
- Failure-Mode Checklist: The four primary ways AI agents fail in the wild, along with actionable fixes for each.
Target Audience
- Founders and technical co-founders building AI into a product and establishing reliable agent-code boundaries.
- Solo builders and small engineering teams (1–10 people) shipping AI features without a dedicated ML function.
- Product engineers who have built a prototype and need it to run unattended with high reliability.
- Startup CTOs and engineering leads evaluating agentic architecture before committing to production.
- Operators and technical entrepreneurs looking to automate real workflows (support triage, onboarding, document processing).
Prerequisites
- Runtime: Node.js 20+ or Python 3.10+ installed.
- LLM Access: An LLM API key from any provider (free or trial tier is sufficient).
- Setup: Starter repository cloned with setup script run (link circulated prior to the session).
- Tools: A code editor of your choice (Visual Studio Code recommended).
- Assumed Knowledge: Comfortable reading and modifying code in at least one language; basic command-line usage. No machine learning or data science background required.