// SoftServe AI Week 2026 ยท October 2
AI lowered the floor, our work now is the ceiling
AI made the first draft free, and most people read that as the great equalizer. This talk argues the opposite: when a rough version costs five minutes instead of five days, the only thing left that separates people is judgment. What to hand off and what to keep, the instinct for when something is quietly wrong, and the concrete decisions that separate a demo from something that holds up with real users: guardrails, evals, cost per inference, and knowing when not to use an agent at all.
Download the slides (PDF, 6 MB) ↓Articles from the talk
- The finished illusion →
- Guardrails: a rule the AI can ignore is a suggestion →
- 108 emails: what an eval is and why to use real cases →
- A date doesn't need an agent, it needs a calendar →
- 163 rules for working with AI agents →
Resources
Guardrails and evals
The lethal trifecta Simon Willison When an agent becomes dangerous to connect to anything. simonwillison.net → Your AI product needs evals Hamel Husain How to build evals from your real data. hamel.dev → AI Incident Database incidentdatabase.ai Real AI failures, including Chevrolet and Taco Bell. incidentdatabase.ai →
Building agents, and where to start
Building effective agents Anthropic Agents vs workflows, and why to start simple. anthropic.com → Skills, including grill-me Matt Pocock Free. Makes the AI interview you before it builds. github.com → Patterns for LLM systems Eugene Yan Evals, guardrails, caching and cost in one read. eugeneyan.com → DotCSV Carlos Santana AI explained in Spanish. youtube.com →
What did you think of the talk?
Your feedback helps me make the next one better.
Leave feedback