AI Agents in Practice - Beyond the Prototype (slides and video from my AI Fokus talk)
In this article
Making an impressive prototype is easy. But what about the icky stuff that comes after - real-life deployment, scaling, monitoring, optimization, data architecture, safety?
On May 19 I gave a talk at AI Fokus 2026 in Stockholm, titled “AI Agents in Practice - Beyond the Prototype”, using real-life examples from agents on the Abundly platform.
Key points from the talk:
- Team-based agents > personal assistants. Abundly agents are designed to work as team members, rather than personal assistants. They work in the same tools you do — Slack, GitHub, Notion — and communicates with other team members and not just you.
- Impressive ≠ useful. Building a cool agent demo is easy now. Making one that's actually useful day-to-day takes more time and persistence than you'd expect — but it pays off big time.
- Finding golden use cases takes experimentation. Look for tasks that waste your time, take a lot of effort, and don't require much creativity. That's where agents shine. For every home run, expect a bunch of duds.
- The post-prototype blues are real. After the initial wow, you sometimes hit a valley where the agent makes mistakes or costs too much. Don't give up — keep iterating. Agents aren't built, they're grown.
- Think of your agent as a super intern. It needs clear context, the right tools, feedback, and supervision — at least in the beginning. When it messes up, don't blame it. Figure out what context it was missing.
- Context management is the #1 optimization lever. Not too little (it guesses), not too much (it gets confused). Use progressive disclosure — let the agent fetch what it needs, when it needs it.
- Agents can self-improve, with guardrails. Let them update themselves based on feedback, but not autonomously. Predictability matters.
- Safety is a design trade-off. Broader scope and more tools = more power but more risk. Use hard guardrails (whitelists, approval flows) instead of relying on the LLM to remember instructions.
Here are some sample slides:
