Descriptions:
Aman Singla presented MarcoPolo at AI Dev 26 in San Francisco — a persistent workspace layer designed to make it practical for non-developer roles like RevOps and support operations to use agentic AI tools against real enterprise data sources including Salesforce, Jira, databases, and data warehouses. The central thesis is that MCP servers and APIs alone are insufficient for reliable production use; agents also need a structured environment with compute, context, and persistent memory.
The architecture embeds DuckDB alongside the LLM so the agent has a real execution environment for loading, joining, and querying data across multiple sources — directly solving the context-explosion problem that occurs when agents attempt to pull multi-table query results into their context window. MarcoPolo also tackles the cold-start problem by curating schema and syntax information for each connected data source at connection time, so agents produce useful queries immediately rather than hallucinating their way through unfamiliar APIs.
Over time, each workspace accumulates query history and cross-source relationship knowledge that improves future queries and creates an auditable record of what data was accessed and when. These workspace knowledge bases can seed new workspaces, functioning like structured onboarding documentation for a new employee or agent. Singla positions MarcoPolo as connective tissue between the growing ecosystem of agentic tools — Claude, ChatGPT, Perplexity, Cursor — and the structured enterprise data those tools currently struggle to use reliably outside of developer-controlled environments.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Showcase
