AI Superstream: Context Engineering

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Introduction – Angelina Yang (5 minutes)

Angelina welcomes you to the AI Superstream.

Keynote – Scouting the Web: What Happens When the Entire Web is the Agent’s Context? – Dhruv Batra (15 minutes)

Imagine an AI agent that monitors the web—for news, product price drops, reservations, tickets, leads—anything you care about. The agent must be always-on for weeks or months, broad in its coverage of the entire web, adaptable to changing information online and users’ preferences, precise in its reporting with citations, self-aware enough to not repeat itself and to contextualize new findings, all while not blowing up its context and your budget. How would you build such an agent? Dhruv Batra, Yutori's cofounder and chief scientist, describes how the company did it.

How Long Contexts Fail (and How to Fix Them) – Drew Breunig (30 minutes)

Million-token context windows promised a new era for AI agents—just throw everything in the prompt and let the model handle it. But in practice, overloaded contexts fail in predictable ways: hallucinations compound over time, models repeat past actions instead of reasoning forward, irrelevant tools degrade output quality, and accumulated information contradicts itself. Writer and technology leader Drew Breunig draws from recent research to show you real examples of each of these failure modes and introduces techniques that keep contexts under control, including selective retrieval, dynamic tool loading, context quarantine, pruning, summarization, and offloading.

Break (5 minutes)

Context Rot: Its Implications and Potential Solutions – Jeff Huber (30 minutes)

We assume that large language models process context uniformly—that is, the model should handle thousands of tokens just as reliably as it does a hundred. But in practice, this assumption does not hold, even on simple tasks. CEO and cofounder of Chroma, Jeff Huber, shares the results of Chroma’s evaluation of 18 large language models, including the state-of-the-art GPT-4.1, Claude 4, Gemini 2.5, and Qwen3, showing how models’ performances grow increasingly unreliable as input length grows.

How to Tell If Your Agent Used the Right Stuff – Apurva Misra (30 minutes)

Many so-called “agent failures” are actually context failures in disguise. Apurva Misra, founder of Sentick, explains techniques that help you tell whether your agent saw and used the right context, including tracing and attribution, golden datasets for context-aware evaluation, and targeted probes to test retrieval quality. You’ll come away with a practical toolkit for approaching bad answers as context problems and for systematically improving reliability without reaching for a bigger model.

Break (5 minutes)

Linking Memory to Context via Knowledge Graphs and Ontologies – Paul Iusztin (30 minutes)

Even though LLMs boast support for inputs of over 150,000 tokens, performance quickly degrades. Also, “bad” data within your context window only compounds the phenomenon known as “context rot.” You might think that the solution is just to keep relevant data in the context window, but it’s not that simple. Siloed data and data aggregated from private databases, email, video meetings, documentation, and cloud create fragmentation that makes retrieving the correct data difficult. A popular solution is knowledge graphs. Paul Iusztin, bestselling author of LLM Engineer’s Handbook, explores how to keep context windows in check by offloading your knowledge into a long-term memory layer built with knowledge graphs and ontologies. Using graph‑based retrieval techniques (GraphRAG), he also shows how to keep only the data relevant to the task at hand.

Building the Brain: How Predictive Memory Transforms AI Agents – Shawkat Kabbara (30 minutes)

Most AI agents fail in production because they can’t remember or predict context the way human brains do. Shawkat Kabbara, founder and CEO of Papr, demonstrates how to build predictive memory systems inspired by neuroscience—from working memory to episodic recall—using custom schemas and GraphQL. You’ll learn why memory is prediction, not search, and see live demos of brain-inspired architecture that achieves 91% accuracy on Stanford benchmarks while delivering less than 100-millisecond response times for intelligent applications.

Break (5 minutes)

Why AI Gets Confused and How to Fix It: 4 Rules for Context Engineering – Yuzheng Sun (30 minutes)

Building reliable AI agents isn’t just about writing better prompts; it’s about managing the information you feed them. Dr. Yuzheng Sun, founder of Superlinear Academy, breaks context engineering down into four practical principles to stop your AI from hallucinating or getting lost: injecting the right information at the right time, removing noise so the model focuses on the signal, storing knowledge outside the prompt for easy access, and breaking complex tasks into smaller pieces to respect the model’s limits.

Engineering Context Quality by Architecting Agent Memory – Mikiko Bazeley (30 minutes)

Mikiko Bazeley, staff developer advocate at MongoDB, shows how you can use MongoDB Atlas and Voyage AI models to improve context quality and agent performance by designing and implementing agent memory via short-term and long-term persistence.

Closing Remarks – Angelina Yang (5 minutes)

Angelina closes out today’s event.