Infrastructure behind Dust deep-dive agent (blog.dust.tt)

Dust announced Deep Dive, an infrastructure overhaul that turns short, speed-optimized agents into long-running, multi-step investigators that can explore an organization’s entire data estate. Triggered by agent logs showing models inventing filesystem-like paths, Deep Dive maps disparate sources (Slack, Notion, GitHub, Sheets, Snowflake) into a synthetic Unix-like filesystem with five commands (list, find, cat, search, locate_in_tree), adds live warehouse discovery and on-the-fly SQL, and auto-discovers workspace toolsets so agents require no per-workspace wiring. The result is a general-purpose work agent that can follow leads across docs, databases, and the web over 10–30+ minute investigations rather than single-shot Q&A. Technically, Deep Dive uses a three-agent system (@deep-dive coordinator, @dust-planning strategic reviewer with no data access, and multiple @dust-task workers) to run up to six parallel sub-agents, each with its own fresh context window, coordinated via durable Temporal workflows. It integrates native reasoning models (extended-thinking chains) by redesigning the internal data model to preserve intermediate thought, and solves finite-context limits with tool-output pruning (replacing large outputs with placeholders but keeping call traces) plus offloaded tool-use (treating big outputs as files). Together these innovations expand compute/time horizons, maintain coherence across parallel workstreams, and set a new infrastructure bar for practical, enterprise-grade deep agent workflows.