LangTalks 2026: When Engineers From the Future Gather in One Room
A conference, a great breakfast, 1,000 engineers, and a few talks that changed how I think about AI development.
Yesterday I found myself in a very unusual place. I felt like a beginner at a conference filled with people who were building with AI long before it went mainstream.
LangTalks AI Engineering Conference 2026 had 1,000 attendees and 3,000 applications. The line at the entrance never fully cleared all day.
The logistics
Doors opened at 08:00. Starting a conference before the workday is unusual, but it worked. We got plenty of parking, fast registration, and a surprisingly good breakfast featuring croissants, sandwiches, and other snacks.
When it came time for lunch, giving grab-and-go food boxes to a thousand hungry engineers was a smart strategic call. The lunch also included some really nice desserts, like handmade marshmallows.
The venue was Habima Theatre. The building is a bit worn, but the atmosphere was great. The conference spread across two floors with balconies, making it easy to watch the crowd. AI-generated video intros on the screens felt completely natural for this crowd.
There was one bathroom issue. At peak hours after lunch, the line reached all the way back to the main hall. This is probably more of a venue problem than an organizer mistake.
What surprised me most was learning that the organizers, Lee Twito and Gal Peretz, are engineers who run LangTalks on the side. Pulling off an event of this scale and quality as a side project deserves real respect.
The content
The program covered a lot of ground with consistently good speakers. A few sessions stood out.
Lee Twito & Gal Peretz: Autonomous agents in production
The hosts didn’t just open the event. They did live demos of the tools they actually use every day, instead of showing mockups or concept slides.
They started with Andrej Karpathy’s quote on Software 3.0 about how programming is turning into prompting. Then they showed what that actually looks like in practice.
Lee demoed PepperClaw, an AI chief of staff running on WhatsApp. Its workflows are defined as markdown files. The system uses a fail-driven loop where the agent rewrites its own rules based on your edits. The feedback loop is baked into the architecture, so it genuinely gets better at understanding you over time. He showed live examples for morning prep, meeting analysis, and drafting follow-ups.
Gal demoed Hyped, a Telegram-based tool for managing virtual developers from your phone. It handles voice inputs, visual planning, and even live calls with your agents.
My main takeaway here was that managing agents needs to look like managing a team, rather than running a script. If your agents are supposed to work like a team, you have to build their communication infrastructure the same way.
Ido Salomon: Agent orchestration and AgentCraft
This was one of the most honest talks of the day.
His first slide just said: “We are the bottleneck.”
We keep spinning up more agents and parallel tasks, but throughput doesn’t scale. The human in the loop is the constraint because there is too much context to hold in your head and too many agents to monitor.
Ido’s solution borrows from real-time strategy games. RTS interfaces were built to manage many asynchronous units with minimal cognitive overhead. AgentCraft applies this exact mental model to AI agent orchestration.
He demoed a game-board-style UI with tabs for Quests, Campaigns, Traces, Analytics, and Memory. Agents are treated like heroes and resources like buildings. You get situational awareness at a glance instead of parsing walls of logs.
He laid out his roadmap as a fantasy map with three goals:
- Visibility: seeing what agents are doing in real time.
- Autonomy: letting agents execute without micromanagement.
- Collaboration: humans and agents operating in the same shared context.
Most tools today are stuck somewhere between visibility and autonomy. Collaboration remains mostly unsolved. AgentCraft bets that the interface metaphor is just as important as the underlying architecture.
Try it: getagentcraft.com
Yoni Levin: Two years of GraphRAG mistakes
I caught this one on video. Yoni delivered exactly what you’d expect: zero filler and real production pain.
His abstract opens with a great line: “Everyone says ‘use a Knowledge Graph for RAG,’ but nobody tells you how much it hurts when you actually try to build it.”
He spent thousands of hours trying to turn complex books into interactive learning experiences using Neo4j and LLMs. He unpacked three specific failure modes:
1. The knowledge extraction problem. Extracting structured knowledge from a 300-page book is fundamentally harder than searching a massive database. Books have narrative, ambiguity, and implicit context, which don’t map cleanly to a graph.
2. The complexity trap. They started with a graph that was too simple to support meaningful queries. As they added complexity, it eventually became too complex to query at all. The query engineering effort negated the benefits of having the graph.
3. RAG vs. GraphRAG trade-offs. He found cases where plain RAG failed and GraphRAG succeeded, like doing multi-hop reasoning across chapters. But GraphRAG also introduced retrieval artifacts that standard vector search would have avoided.
GraphRAG is a powerful pattern, but his timeline is a good reality check. You have to start simple and learn what “too complex to query” looks like.
Read more: medium.com/@yonatanvlevin
Why a QA Architect goes to an AI conference
I call myself a Quality Entrepreneur, so what was I doing at a conference for data scientists and ML engineers?
Whenever I design engineering solutions, I try to understand the raw materials first. You wouldn’t build a house without knowing the local building materials or tool availability.
Since AI entered the picture, my team at Skipper Soft has been working with a completely new material. It’s a dynamic space, and we’re still figuring it out.
I didn’t go to LangTalks just to get inspired. I went for calibration. I wanted to see where the frontier actually is, what problems are already solved, and which tools actually work in production rather than just in a demo.
My main takeaways from the day:
We are shifting from writing code to managing agents. This feels like a structural shift in engineering. Tools like AgentCraft, PepperClaw, and Hyped are early signals of how our workspaces will look in a few years.
Honest post-mortems beat success stories. Hearing about Yoni Levin’s GraphRAG mistakes was much more valuable than listening to a generic success story. Learning from failures is a real advantage right now.
Scaffolding determines autonomy. When you build the right environment, agent autonomy starts to compound. It’s an architectural property of the system.
Final thoughts
I never go to conferences just for the talks. Meeting the people who are actually building this stuff provides value you can’t get from a YouTube recording. LangTalks drew a crowd with concrete experience, and a few hallway conversations gave me more insight than most online courses.
It also helps me update my mental map of where the industry is heading and figure out my place in it.
The conference answered some questions and raised new ones that I’ll need to dig into.
If you are building AI tools, keep an eye on LangTalks: langtalks.ai. It is one of the few places where people focus on what actually works.
Attended LangTalks AI Engineering Conference 2026, Tel Aviv, May 4, 2026. Happy to discuss in the comments, especially if you were there or are working with GraphRAG or agent orchestration.
Follow if you’re interested in AI tooling for QA and Quality Engineering.