Agentic Trick or Treat: AI Horror Stories & Lessons Learned

Agentic Trick or Treat

AI Horror Stories & Lessons LearnedMonday, November 2, 2026 | 18:00 to 20:15Microsoft Reactor, 156 Menachem Begin Road, 50th Floor, Tel AvivThe talks will be held in Hebrew.

Working with AI is full of surprises, and not all of them are the good kind. Sometimes the scariest part is finding out what it actually did.Halloween lands on a Saturday this year, so we are holding it on Monday, which is when most scary things get discovered anyway. Join us for an evening of AI horror stories from teams at Tikal, Microsoft, Port, JFrog and Qodo: what went wrong, what they changed, and what you can apply on your own team the next morning. We close with a panel on the AI Software Factory: can we trust the process?Agenda

18:00 to 18:30 | Welcome and NetworkingMeet other engineers and the speakers.

18:30 to 18:45 | Opening RemarksA few words to kick off the evening.

18:45 to 19:45 | AI Horror StoriesFour talks, 15 minutes each.--------

We automated the assembly line, then the inspector, then the inspector's inspector.Erik Zaadi, R&D Director, PortWhen agents multiply the output, the reviewer becomes the bottleneck, and could potentially cause rubber stamps.Our agents were producing more PRs than humans could meaningfully review, many of them technically correct and entirely beside the point. At the same time, the bots we built for other teams added to the pile: support agents correlating tickets and opening bugs faster than engineering could read them. Nothing was broken, everything was busy, and nobody could tell which of it mattered.Erik will share how the noise built up, why our first response was (of course) to add more AI, what that did and didn't fix, and what we measure and no longer trust today.

I See Dead Products: Building AI Assistants at ScaleNoa Barbiro, Principal PM Manager, Memory & Search, Microsoft Copilot

Agents are new. The ways they fail aren't.Years before ChatGPT, Booking.com's Tel Aviv team shipped a voice assistant to millions of travelers. Noa will share an honest take on the AI horror stories that came with it: five lessons from being early to market with an AI product, and how they shaped Copilot Memory in Microsoft 365 Copilot. You'll leave with practical patterns for building assistants and agents that follow user intent, catch failures before your users do, and earn the right to remember.

The Curse of the Query RewriteSigal Shaharabani, Backend Tech & Group Lead, TikalUsing AI to migrate SQL queries sounds straightforward, until differences in data types, functions and execution behavior make a convincing rewrite behave very differently.Sigal's team set out to automate a migration from an RDBMS to a columnar data warehouse with AI agents. The first attempts failed, and so did several that followed. She will explain why, what context and validation the agents were missing, how the workflow was redesigned to produce reliable results, and where the agents did help.

The Disappearing Token BudgetLiran Baba, AI Lead, JFrogA token budget meant to last 30 days was gone in less than a week. Liran will show where the tokens actually went, what it cost, and the changes that cut unnecessary usage and made their agents more efficient. He will also share what they would do differently next time.

19:45 to 20:15 | Panel and Audience Q&A

AI Software Factory: Can We Trust the Process?What happens when AI becomes part of the whole software development lifecycle? Where does the process break as adoption grows, who reviews the output, and how do teams know faster delivery is not costing them quality?

Moderator: Batel Zohar Tova, Developer Advocate and Azure Senior Technical Specialist, Microsoft

Panelists:Amit Svarzenberg, CTO at Microsoft for StartupsLior Kanfi, Founder and CEO, TikalErik Zaadi, R&D Director, PortLiran Baba, AI Lead, JFrogGuy Vago, Senior Solutions Architect, QodoPlease NoteRegistration is subject to approval by the organizing team. Your place is confirmed only once you receive approval.