LangTalks AI Engineering Conference 2026

As we run more agents in parallel, it becomes clear: we are the bottleneck. Luckily, the skills we need for effective multi-agent orchestration aren’t entirely new, they’ve just been hiding in unexpected places.Through AgentCraft, the game-inspired agent orchestrator, I’ll explore how we can raise the ceiling of human-agent collaboration without burning out in the process.

Gal and Lee, hosts of the LangTalks podcast, will break down what the next generation of autonomous agents will look like and how we can interact with them effectively.

Building the #1 deep research agent was not about adding more complexity, but about doing less with focus and discipline. In this talk, we share our journey from early experiments to production, the core challenges we faced, how we approached token efficiency and evaluations, and what the next generation of research agents will look like.

Silent failures are the hardest bugs in AI engineering: the system "works," but it's quietly suboptimal: less accurate, less reliable, or more expensive than it should be. What are the most common pitfalls that lead to this hidden underperformance, and how can we design systems to detect it and prevent it early?

Creating agents that live and work inside a discord channel and talk to each other. We can add a new agent and show how they work together and can be modified.

Building a cool AI demo is easy; maintaining hundreds of unique deployments for healthcare clients without ballooning your engineering head-count is the real challenge. At Hyro, we spent five years evolving from a code-forward NLP approach to a sophisticated, LLM-driven multi-agent system handling over 700,000 conversations weekly. We share the honest growing pains moving to a unified, scalable platform where Solutions Engineers can deploy complex, reliable agents independently.

In early 2026, our PM started opening pull requests to our production codebase. Not prototypes — real, merged code. This talk is about what that demanded from the engineering team, and from the PM herself. AI didn't eliminate our bottleneck — it moved it. Code generation stopped being the constraint. Review, coordination, and architectural alignment became the new pressure points. Merge rate — the percentage of PRs that actually land in production — became the signal that told us whether the team was adapting or drowning. Getting there required change on both sides. Engineers had to rethink how they review, what they gate on, and how much bandwidth they allocate. The PM had to learn which changes she could ship independently and which required coordination first. Neither side got it right immediately. Once we found that alignment, the effect compounded. Features that used to take a full sprint now land in days. Entire product areas are owned end-to-end by the PM. The developers freed from feature delivery work at a higher level — focused on architecture, patterns, and system design. The team didn't just absorb the change. It got faster because of it. I'll share the real data, the process breakdowns, and what we changed from the CTO's chair.

Autonomous BizDev AI employee for SLED AI. Finds, validates and acts on the US government market.

In the era of coding agents, scaling a codebase isn't just about writing more code faster — it's about building systems that agents and humans can navigate safely and predictably. Live coding demo covering: the Dev-X setup built for both developers and coding agents, Skills & Standards, and why SOLID principles and Design Patterns matter more than ever when agents are writing your code.

Cloud platforms generate massive security data: events, resources, connections. Manual investigation doesn't scale. We built fully autonomous AI agents that investigate from alert to verdict, operating 10x faster than humans through context management, adaptive replanning, and evaluation metrics.

Everyone says "use a Knowledge Graph for RAG," but nobody tells you how much it hurts when you actually try to build it. After thousands of hours trying to turn complex books and multi-modal content into an interactive learning experience, we share every mistake made with Neo4j and LLMs — why extracting knowledge from a 300-page book is harder than searching a billion-row database, how the graph evolved from "too simple to work" to "too complex to query," and real-world examples of where RAG failed and how GraphRAG saved (and occasionally broke) the experience.