AI
What an AI Agent Actually Is
No single definition of an AI agent exists. Kniberg’s version stresses autonomy: an AI agent is a digital entity that uses an LLM as its brain, holds a mission beyond answering chat messages, and has tools to act on the world. It can react to events, run on a timer, and do things on its own accord.
The useful mental model is an intern with superpowers and strange limitations. That framing matters because it sets the right expectations: the agent needs a job description, a place to live, tools to do the work, and someone to answer to when things go sideways.
Building an Agent Live on Stage
Kniberg built Mrs. GoTo live on stage using Claude 4.5. In the Mrs. GoTo GitHub and Slack demo (03:52) the agent requested its own tools, scanned 30 commits, identified the active repo, and posted a categorized dev update to Slack. Then, in Mrs. GoTo’s phone call in Danish (07:16) she called Henrik’s phone, delivered the Git summary in Danish, and refused to switch languages. After the demo she scheduled a weekly Friday digest, set up email-triggered support, and wrote her own job description.
Picking the Right Work to Hand Off
Agents sit between code and humans. Code is fast and perfectly predictable but has no intelligence. Humans are intelligent but slow and inconsistent. Agents are more consistent than humans (no Friday-afternoon fatigue) and more capable of reasoning than code, while being slower and less reliable than either at scale.
Pick tasks that take a lot of time, feel like a waste, and need only moderate judgment. Research, classification, compliance review, routing, and scheduling all fit. Agents get lost on genuinely complex open-ended work, so start narrow.
What Good Agent Use Cases Look Like
An investment firm screened thousands of companies for potential investments. Human analysts produced inconsistent results and hated the work. A well-prompted agent processed every company in parallel, bucketing each into yes, maybe, or no. Domain experts compared its decisions to earlier human calls and found the agent was more often correct. Not because it was smarter, Kniberg said, but because it was more patient and consistent.
A Swedish TV channel’s scheduling agent handled photojournalist shift changes daily, accounting for union rules and 11-hour rest requirements. Understanding that domain took longer than building the agent.
Data Format as Agent Architecture
A 1,400-cell schedule spreadsheet fed as CSV into an LLM context failed at scale. Even Claude Opus hallucinated. The fix: convert to JSON where each record states when, who, what, and where, then give the agent four CRUD functions instead of the full dataset. In the data documents demo (22:41) one prompt created a MongoDB-backed animals database and a browser app for editing it, all reads and writes bypassing the LLM.
“kind of a big insight. That small change.” — Henrik Kniberg
For the TV scheduling case, the agent wrote its own import script and used query tools to answer shift-change questions. Token use dropped. Accuracy improved.
Design Principles, Safety, and Debugging
Give each agent a specific job, a limited tool set, and a named owner. The scope-versus-capability dial runs from a narrow tag-an-incident agent up to broad autonomous work. Broader means better models, tighter prompts, and more monitoring, not automatic danger.
“Don’t start here. Start down there.” — Henrik Kniberg
When an agent misbehaves, look at what context it had. Hallucination is a design flaw, not a feature of LLMs. In the HTTP tool map demo (33:20) Claude 4.5, given only an HTTP tool, chained Perplexity search, OpenStreetMap geocoding, and Leaflet.js to render a live venue map with bus times.
“hallucination just means your agent is designed wrong.” — Henrik Kniberg
Notable Quotes
kind of a big insight. That small change. Henrik Kniberg · ▶ 25:03
Don’t start here. Start down there. Henrik Kniberg · ▶ 30:36
hallucination just means your agent is designed wrong. Henrik Kniberg · ▶ 38:09
Key Takeaways
- Start agents on tasks that are time-consuming, feel like a waste, and require only moderate intelligence.
- Replacing CSV context dumps with JSON plus CRUD tools cuts token use and stops hallucination at scale.
- Hallucination is a design flaw: wrong context or wrong model. Fix the design, not the expectation.