Glossary

Sovereign & agentic AI, defined.

Clear, plain definitions of the terms behind planned execution, the Company Brain, sovereign agentic AI, and the technology Eigen Networks builds.

Task orchestration and execution layer: The structured layer between enterprise applications and the models, tools and data they use. It decides how a task is executed, rather than describing what happened after the fact.
Open agent loop: The common agentic pattern where a model receives a task and a large set of tools, decides what to do next, calls a tool, reads the result and repeats. Context accumulates, the workflow is rediscovered every step, and independent work still runs sequentially.
Task compilation: Turning a business objective into an executable, governed workflow before expensive model execution begins. Not planning in the conversational sense: the output is a plan the execution layer runs, validates and records.
Execution plan: The dependency structure of a task: what has to be accomplished, which steps depend on others, which run in parallel, which information and tools each step is permitted, which steps need model intelligence, which run deterministically, and where validation or approval applies.
Execution receipt: The record returned with an outcome: what was planned, what was executed, which evidence each step used, which models and tools contributed, which validations and policies applied, what it cost, and why the final action was selected.
Cost per successful task: The commercial metric for an execution layer. It measures outcome and cost together, because a cheaper execution is worth nothing if it produces a worse result.
Compatibility Mode: Integrating an execution layer by changing the API base URL of an OpenAI compatible client. It produces a live measurement and the optimizations that are safe without owning the whole agent loop.
Execution Mode: Integrating through a task and tools contract: the complete task, the available tool or MCP contracts, the relevant policies, the required output schema and the configured model route. The execution layer then owns the execution loop.
Procedural knowledge: Knowledge about how work is performed rather than documents about it: how objectives are decomposed, which systems hold authoritative evidence, which tools suit which operation, which policies control a decision, and which execution patterns reliably succeed.
Sovereign AI: AI where the models, the data and the infrastructure all stay under the organization's own control. Nothing is processed or retained by a third party outside the organization's perimeter.
Agentic AI: AI that does not just answer questions but acts. Agentic systems plan, make decisions, collaborate and carry tasks through to completion.
Sovereign Agentic AI: Autonomous, multi agent AI that runs entirely inside an organization's own infrastructure, so agents plan and act on the company's own data with no data leaving its perimeter.
Agentic AI Network: A coordinated network of specialized AI agents that plan, collaborate and execute together, orchestrating many small models rather than relying on a single large one.
Company Brain: The reusable, versioned procedural knowledge that accumulates from validated executions: plans, tool relationships, evidence paths, policies, exceptions, approvals and outcomes. A structured representation of how a company's work is actually performed, not a language model or a document search system.
Small Language Model (SLM): A compact language model small enough to run privately and economically on an organization's own infrastructure. Working together as a network, SLMs deliver enterprise grade accuracy without the cost or exposure of frontier APIs.
Data Sovereignty: The principle that an organization's data stays subject to its own control and to the laws of its own jurisdiction. In AI, it means data is never sent to or processed by external services.
On-premise AI: AI that runs inside an organization's own data center or environment rather than in an external cloud, so data and models never leave the organization's control.
Air-gapped AI: AI that runs in an environment with no outbound network connectivity at all. Essential for classified, defense and highly regulated settings.
Zero data egress: A property of a system where no data leaves the organization's perimeter. Inputs, outputs and intermediate results all stay inside the organization's own infrastructure.
Human in the loop: A design where AI agents act autonomously when confident and escalate critical or uncertain decisions to people at the right moment.
Model-agnostic AI: An architecture that is not tied to any single model or vendor, so the underlying models can be swapped and evolved over time without re building the system.