Exa's Jon details how their company maintains a massive, continuously updated semantic index of 100 billion pages to serve AI agents. The talk covers Exa's unique incremental indexing system, a custom data processing framework called "Skeleton" that handles 100,000 pages per second, and "Pythia," an elastic, cost-aware GPU orchestration system for managing both real-time ingestion and large-scale index re-embeddings.
Gil Feig, co-founder and CTO of Merge, explains why agents need a company-wide context graph rather than a few MCP connections. Using the question of which customers were upset last week, he shows why live API lookups fail across large datasets and why synced local copies, routing, summarization, skills, memory, and traceable provenance are needed to build a reliable company brain.
Jakub Hojsan, a forward deployed engineer at Exa, explains how search built for coding and code-review agents closes the gap left by model knowledge cutoffs. He covers why agents need explicit rules for when to search, how query-dependent highlights cut a 100,000-character page to about 500 characters without added latency, and how Exa compares with native web search on transparency, cost, and model-provider independence, with a preview of Exa Agent.
Dylan Couzon, developer relations engineer at Qdrant, argues that frontier-class models now run on hardware people can own, but the memory that makes an assistant personal still lives in someone else's cloud. He covers why owning inference gives autonomy while owning memory gives continuity, the CPU, RAM, and disk model for LLMs, why retrieval beats stuffing everything into the prompt, and a live offline demo of a drone building searchable memory with Qdrant Edge, plus opt-in shared memory.
Roman Ugarte, a key figure behind Grok Bot, shares insights into its rapid development, emphasizing the strategic choice to build it as a separate, cloud-first product with bots having their own dedicated 'computers.' The discussion highlights a "colleague-pilled" product philosophy focused on full delegation and constant reinvention, enabling the creation of autonomous AI teammates for both work and personal life.
Experiments with an AI tutor reveal that not compacting context often outperforms summarization, leading to better recall, lower cost, and faster responses, primarily due to LLM API prompt caching. Compaction invalidates the cache, making it unexpectedly expensive. While cloud environments benefit from full history, local hardware limits necessitate compaction, highlighting the need to prioritize actual constraints over default compaction strategies.