The 'First-Day Hire' Problem

Edo Liberty argues that current enterprise AI agents suffer from a fundamental failure in 'theory of mind.' While agents are brilliant, they lack the context of a seasoned employee. They start every session as if it were their first day, requiring them to read through massive amounts of documentation to understand company culture, processes, and tribal knowledge. This results in agents that are slow, expensive, and often inaccurate because they lack the 'sum total' of company experience that isn't captured in a single document.

The Knowledge Layer Architecture

To solve this, Liberty proposes a persistent, specialized knowledge layer (Nexus) that exists independently of the agent. This layer is defined by three core components:

  • The Manifest: A meta-knowledge file that defines what the agent cares about, including specific tasks, entities to track, and guiding principles. This replaces the need for manual, hard-coded skills or plugins.
  • Contextual Storage: A multi-modal storage system that includes semantic maps (for schema and glossary), markdown files (for unstructured memory), SQL tables (for facts), and graph databases (for causal relationships and entity mapping).
  • Curated Updates: The system automatically assimilates new information (PDFs, meeting transcripts, etc.) by updating the graph, tables, and indexes based on the manifest, ensuring the knowledge remains current—a critical requirement for enterprise data.

Runtime Coding Agents vs. Prompt-Heavy Tooling

One of the most significant technical shifts proposed is the move from massive prompt-based tooling to runtime coding agents. Instead of providing an agent with 150,000 tokens of documentation and instructions, the system acts as a Python REPL (Read-Eval-Print Loop).

When a query is made, the agent writes and executes code to retrieve the answer, iterating until the result is correct. This approach:

  • Reduces Token Consumption: Cuts tooling prompts from ~150,000 tokens to under 1,000.
  • Increases Flexibility: The answer is derived through executable logic rather than a static database query.
  • Improves Performance: Early access results show 77-90% reductions in token consumption and 20-77% faster response times, while simultaneously increasing accuracy by providing the agent with a grounded, structured view of company data.