Glossary
An agentic system is an LLM with hands. The model still generates tokens, but a runtime around it reads those tokens as plans, executes the plans by calling tools (HTTP APIs, databases, the file system, other agents), feeds the results back into the next reasoning step, and persists state across the chain. The compliance burden grows with every tool the agent can touch.
Back to glossaryA single-prompt LLM has one input, one output, no memory beyond the context window, and no side effects. An agent has open-ended input, multi-step output, persistent state, and side effects on external systems. The implication is that 'evaluating an LLM' (latency, hallucination rate) does not evaluate an agent — you also have to bound which tools it can call, which data it can read, and which actions it can commit.
Agentic systems open three failure modes traditional ML governance does not handle. First, decision opacity: the chain of intermediate model calls is rarely captured. Second, side-effect drift: a tool call mid-chain can move money, modify a database, or send a message — outside the original audit trail. Third, jurisdictional leak: an agent that invokes an embedding API or retrieval service in another region quietly defeats the residency posture of the originating system.
These gaps are why per-call cryptographic evidence (Trust Receipts) and supply-chain manifests (AgentBOM) are now arriving as named primitives.
A production-ready agent fixes those three modes by construction. Every step is logged with its inputs, outputs, model, region, and policy decisions. Every tool call returns through a signed evidence layer. Every chain has a stable execution id that cross-links the per-step receipts into one audit story. AgentAnywhere Sovereign's posture treats this as default; it is the difference between an agent demo and an agent a regulator will let into a regulated workload.
Where the regulatory or technical authority for this term actually lives. We cite primary sources so this entry can be checked, not just trusted.
Last reviewed: .
We maintain canonical definitions for sovereign AI, Trust Receipts, data residency, AgentBOM, and agentic AI so procurement, security, and legal teams can quote a primary source instead of paraphrasing one. Email enterprise@soverai.ai if you need an extended PDF reference for a specific regulator.