Enterprise agents need a persistent knowledge layer for company goals, processes, and tribal knowledge.
2
A knowledge layer must stay current so new facts can override older information spread across many documents.
3
Nexus uses a runtime coding agent that writes and executes code at query time, reducing the tooling prompt from about 150,000 tokens to less than 1,000.
Summary
Edo Liberty argues that current enterprise agents have a theory-of-mind problem. They may be capable and well connected to tools, but they do not know what a company knows, how it works, or which facts have priority. He divides knowledge into general knowledge, specific knowledge such as contracts and code, and tribal knowledge about processes, ownership, culture, and decisions. Search and retrieval handle specific knowledge reasonably well, while tribal knowledge is spread across many sources. Liberty presents Nexus as a persistent, specialized knowledge layer shared by agents. It uses manifests to define what a domain should track, then maintains semantic maps, documents, SQL tables, vector search, text search, and graphs. Its query system lets an agent choose how much time or money to spend on an answer. A runtime coding agent writes and runs code while answering, which Liberty says reduced the tooling prompt from roughly 150,000 tokens to under 1,000.
Enterprise agents are brilliant new hires who start every session from zero
Liberty compares an enterprise agent to a brilliant new hire. The agent has tools and high expectations placed on it, but it does not know the company's goals, culture, priorities, or processes. It reads documents and starts most tasks from scratch. Each new session is like the employee's first day on the job. Liberty says the answer is not simply a smarter model or more tools. Agents need a persistent knowledge layer that gives them a continuing understanding of the company.
Company knowledge has general, specific, and tribal parts
General knowledge is information available in the public domain and absorbed during model training, such as improving Rust code or explaining legal terms. Specific knowledge is private information about a company's codebase or a customer's contract. Retrieval, search, and vector databases already address much of this category. Tribal knowledge is different. It includes how work gets done, who owns what, which processes matter, and what principles guide decisions. It is distributed across many sources rather than stored in one searchable document.
A knowledge layer must persist, specialize, and stay current
Liberty says the layer should be built once, or at least rarely, and made available to every agent in the company. It should also be specialized by domain. HR, kernel optimization, and customer support agents need different information, with domain experts able to maintain what each area contains. The hardest requirement is currency. If a company replaces its CEO, employees know the new fact immediately even though thousands of older documents still name the previous CEO. The current fact has to override those older sources.
The knowledge layer supplements existing agent tools
Liberty does not propose replacing language models, local files, search, retrieval, or vector search. Nexus is a separate data source that all the agents can use. Its connectors bring information into the system while respecting access rights. The layer then organizes company knowledge into specialized contexts for areas such as HR, operations, and engineering. This separation reflects Liberty's view that company context has different properties from ordinary documents or public knowledge.
Manifests tell Nexus what a domain should remember
A context in Nexus includes the assets and information associated with a topic. Its manifest describes the tasks to complete, the entities to track, and the information that matters. The knowledge layer uses that specification to organize and maintain its own data instead of relying on engineers to write skills, plugins, commands, or Markdown files by hand. The context can include a semantic map, Markdown records, SQL tables, vector and text indexes, and graphs of entities, relationships, and causal chains.
Import and curate turn new material into maintained context
Import brings new files, events, or concepts into the system. Curate takes material such as a PDF, meeting transcript, or board presentation and updates the system's view of the world. It uses the manifest to decide which concepts to track and the semantic map to locate the relevant tables, relationships, graphs, and indexes. Search provides an internal way to access this information without requiring an agent to understand how each part of the context is stored.
Agents can set a budget for how much investigation a query receives
Nexus uses NoQL to specify how information should be retrieved from a context. Liberty says an agent should be able to provide a token or dollar budget. A small budget can produce a quick, immediate answer, while a larger one allows the system to inspect more of what it knows. The result is grounded, structured text intended for another agent rather than a chat interface.
Runtime coding answers questions by writing and executing code
Nexus uses what Liberty calls a runtime coding agent. At query time, it builds something like a Jupyter notebook or Python REPL, writes code, executes it, checks the result, and rewrites parts that do not work. The final code completes the task and can be saved or rerun. This approach produces answers through flexible code instead of a fixed database query. Liberty says it also reduces the amount of prompting needed to describe the available tools.
"We went down from having something like 150,000 tokens to give our agents all the tooling they need to less than a thousand tokens to specify all the interfaces they need to get information out of Nexus."16:50
Who should watch
You are building agents for a company and keep having to explain its processes, ownership, or priorities in every session.
Your retrieval system finds relevant documents but cannot decide which current company fact should override older material.
You want to understand how a runtime coding agent can replace a large collection of hand-written tool descriptions.
