Glossary: agentic, RAG, environments
This family groups the notions that describe where and how interpretive governance applies. A single concept (canon, authority, evidence) does not deploy in the same way depending on whether a system is agentic or not, operates on the open web or in a closed environment, and relies (or not) on a RAG pipeline.
Each entry links to: a canonical definition, a framework (if applicable), and related pages for applying governance on the correct surface.
Quick access
Terms in the “agentic, RAG, environments” family
Agentic
A system is called agentic when it goes beyond merely responding: it acts (tools, navigation, writing, execution, transactions) and produces effects in the digital world.
- Definition: Agentic
- Related page: Agentic: governing AI that acts
- Framework: Interpretive governance for AI agents (open web & closed environments)
Non-agentic systems
Generative systems that produce outputs without executing tooled actions (or with effects limited to the response), which modifies the risks and evidentiary obligations.
- Definition: Non-agentic systems
- Clarification: Non-agentic systems and interpretive governance
Open web vs closed environments
Two distinct surfaces: on the open web, truth is subject to neighborhood and exogenous signals; in a closed environment, truth depends on internal governance (corpus, access, evidence, journals).
- Framework: Interpretive governance for AI agents (open web & closed environments)
- Framework: Agentic risk matrix (open web & closed environments)
RAG (Retrieval-Augmented Generation)
Architecture where the response is conditioned by a retrieval. The main risk becomes: what is retrieved, how it is interpreted, and when a non-response is mandatory.
Enforceable response conditions
Explicit constraints that determine when a response is authorized (minimum evidence, inference perimeter, legitimate non-response). In agentic systems, these conditions become a security measure.
Identifier governance (disambiguation)
Mechanisms to prevent entity confusion (similar names, homonyms, mixed attributes), particularly when an agent combines multiple sources in a single decision cycle.
- Framework: Identifier governance: entity disambiguation
- Related definition: AI disambiguation
CTIC (cross-layer transactional coherence)
Extension targeting the coherence of dynamic variables (price, stock, delivery, promotions) between layers (canon, retrieval, response, execution), critical as soon as an agent acts or recommends.
Related frameworks and pages (recommended)
Previous page: Glossary: capture, contamination, collisions
Next page: Glossary: sustainability, debt, correction
Phase 7 expansion: retrieval governance and documentary chains
The RAG entry in this family is now expanded into a canonical retrieval-governance layer. Start with RAG governance, then follow retrieval control, source admission, corpus admissibility, retrieval provenance, chunk authority, documentary chain, correction budget, and resorption.
Dedicated family page: Glossary: RAG, retrieval, and documentary chain.
Phase 8 expansion: agentic execution and transactional control
The agentic entry in this family is now expanded into a canonical execution-control layer. Start with agentic risk, then follow multi-agent chains, delegated action, tool-mediated authority, execution boundary, transactional coherence, cross-layer transactional coherence, and agentic response conditions.
Dedicated family page: Glossary: agentic execution and transactional control.
Phase 9 routing layer: memory, persistence, remanence, and correction
This page now routes stateful interpretation questions toward the phase 9 canonical layer: memory governance, agentic memory, memory object, persistent assumptions, controlled forgetting, stale-state handling, surviving authority, interpretive remanence, interpretive inertia, version power, state drift, and correction resorption.
The routing rule is direct: do not infer current authority from persistence alone. A memory object, old citation, surviving source, retrieved fragment, or previous answer must pass freshness, authority, traceability, and correction-resorption checks before it can govern a new response or action.
How to read this lexical family
This family connects retrieval, agents and governed execution. A RAG environment can retrieve documents without becoming agentic. An agentic environment can act without using RAG. The risk becomes more complex when both layers are combined: retrieved material can shape an answer, and the answer can shape a tool-mediated action.
The important distinction is between access and permission. Retrieval gives the system material to consider. It does not automatically give the system authority to conclude, recommend, refuse, execute or delegate. Agentic behavior adds another layer: even a correctly retrieved source may be insufficient to authorize a consequential step.
Typical misreadings
The common error is to say that RAG makes an AI system safe because it grounds responses in documents. That is only partially true. RAG can improve documentary access, but it does not by itself solve source admission, chunk authority, response conditions, non-inference rules, execution boundaries or memory persistence.
A second error is to collapse all agentic failures into hallucination. In an agentic RAG environment, the failure may come from a retrieved source that was obsolete, an admissible source used beyond its authority, a chunk detached from its governing context, a tool call that exceeded the execution boundary or a memory object that survived a correction.
Use in audit and routing
This family should be used when assessing systems that combine retrieval, orchestration, tools, memory or multi-step workflows. The audit should follow the chain from source admission to retrieval, from retrieval to response, from response to delegation, and from delegation to execution.
For routing, this family should support RAG governance, agentic risk, multi-agent audits and execution control pages. It is a bridge family, not a primary definition for interpretive governance as a whole.