Henrik Kniberg shares two and a half years of hands-on experience building AI agents at GOTO 2025. He defines agents as autonomous digital entities using LLMs as their brain, with tools and a mission beyond simple chat. A live demo shows an agent connecting to GitHub and Slack, making phone calls, and scheduling recurring tasks autonomously. Key lessons include: identifying tasks that are time-consuming, low-value, and moderately complex as ideal agent candidates; using structured data formats (JSON with CRUD tools) instead of dumping raw CSVs into LLMs for scalability; having agents write their own code and scripts to avoid token waste; treating agents like interns (useful metaphor for debugging and design); keeping humans in the loop for safety; iterating heavily on prompts; and managing risk by scoping agent jobs narrowly and adding guardrails as complexity grows. A real-world case study of a scheduling agent for a Swedish TV channel illustrates both the potential and the challenges. The talk also previews agents building their own tools via HTTP requests.
Sort: