Dan Kornberg
← All projects

Talks & workshops · New York · November 2025

I headlined a room of 70+ founders and engineers on agentic workflows.

The talk: when one agent isn't enough. How to stop prompting a model and start designing the system around it, with a live build in n8n.

  • Headline talk at AI Agents in Practice | NYC Edition
  • 70+ founders, engineers, and product people in the room
  • Workshops inside Wix on agents, multi-agent systems, my tools, and IDE workflows for non-developers
Dan on stage with a microphone, next to a screen showing a four-agent workflow: search trending topics, draft a blog post, edit it against the style guide, publish it. Live · Nov 2025
AI Agents in Practice | NYC Edition

01 · The talk

One talk, from a single prompt to a system of agents.

The room was founders, engineers, and product people. I walked them from asking one model for one answer to designing a workflow of agents, covered the principles that make those systems work, and then built one live.

AI Agents in Practice · NYCNov 2025
Headline talk

When one agent isn't enough.

Agentic workflows: from prompting a model to designing the system around it.

  1. Context ≠ attention: why a 200k window still misses the needle
  2. Self-refinement: a fresh critic agent, without the author's bias
  3. Specialization: one agent, one job, the best model for each step
  4. Separation of concerns: PII scrubbing, structured output, per-step evals
  5. Modular systems: human-in-the-loop approval, UX beyond chat
  6. Operations: token budgets, decentralized ownership, non-devs who build
Live in n8n: Research→Draft→Edit→Publish
Dan KornbergAI strategist · Wix
70+Founders, engineers, product

A recap poster, made for this page.

02 · The shift

Stop prompting. Start designing systems.

A standard prompt is one instruction and one answer: fast, but fragile. An agentic workflow can research, create, analyze, and iterate, with each step feeding the next.

Standard prompt · Fast but fragile
Agentic workflow · Research, create, analyze, iterate

03 · The principles

Why one agent isn't enough.

“The context window is big enough” and “I ask it to critique its own output” sound fine until a task gets complex. As context fills, attention dilutes, and a model is a poor judge of its own work. So split the job up.

01Context ≠ attention

A model can read 500 pages and still miss the one rule in the middle. Give each agent one goal to focus on.

02Specialization

Researcher, drafter, editor: the best model for each task, only the context it needs, and a system prompt tuned for one job.

03Separation of concerns

An agent that cleans PII before data reaches outside tools. And if agent 2 misses, you fix its prompt without touching agent 1.

04 · Live in n8n

Then I built one, on stage.

A blog post workflow in n8n, on a schedule: a research agent searches the web for trending AI topics, a drafting agent writes against the writing guidelines, an editor agent checks it against the editing guidelines and returns structured output, and the post gets published. It's the same idea behind my UX writing agents at Wix.

The demo workflow.

05 · Inside Wix

At Wix, I taught it in workshops.

The same ideas, for colleagues: what agents are and how they work, how multi-agent systems fit together, how to use the tools I built, and how non-developers can build reusable workflows in an IDE.

  • WorkshopHow agents work
  • WorkshopMulti-agent systems
  • WorkshopThe tools I built
  • WorkshopReusable IDE workflows for non-developers

06 · The takeaway

Start small: automate 10% of one thing.

At Wix I mostly built workflows for content writers and product people, but this helps anyone, in any role. What I told the room, from my own path as a non-technical builder:

  • Building today is less about coding and more about architecting.
  • If you can think of it, you can build a proof of concept.
  • Get your hands dirty today, start small, scale later.
10% of one workflow you touch every day. n8n is an easy way to start.
The takeaway.

More detail

The full outline
  1. The agentic approach: from standard prompting to iterative workflows.
  2. An example: one blog post prompt, split into four agents.
  3. “Is that really necessary?” Where one agent breaks: attention and self-refinement.
  4. A more complex version: scraping sources, outline and SEO docs, style and grammar guides, a publish schedule.
  5. Where to build, and a live n8n demo.
  6. Benefits: better execution, separation of concerns, modular systems, and operations.
  7. The non-technical builder: lessons from my own path.
What's under the hood
  • Agentic workflows
  • Multi-agent
  • Model selection
  • Structured output
  • PII handling
  • Evals
  • Human in the loop
  • Token budgets
  • n8n
Questions a hiring manager might ask
Who was in the room?
A packed room of 70+ founders, engineers, and product professionals at AI Agents in Practice | NYC Edition, in November 2025. Dan gave the headline talk.
Where did the material come from?
His own work. At Wix he mostly built workflows for content writers and product people, like the UX writing agents and the glossary check.
What did the workshops at Wix cover?
How agents work, multi-agent systems, the tools he built, and building reusable workflows in IDEs as a non-developer.