Analyst vocabulary, mapped honestly

If you evaluate AI tooling, you have met these words: agent sprawl, guardian agents, AI TRiSM, discover / observe / govern / secure. They are useful shorthand, and a 2026 buyer expects a vendor to speak them. They are also easy to abuse — to imply a product is a category when it merely sits near one.

This page does three things: it defines each term, it attributes it where it has a real owner, and it says plainly which ones describe Olivares AI and which ones we only relate to. For the numbers that back the underlying market, see Market context & sources.

Agent sprawl

What it means. The uncontrolled proliferation of AI agents, copilots, MCP servers and automations across an organization — created by different teams, with different credentials, touching different systems, faster than anyone keeps an inventory. The result is unknown agents with unknown access.

Does it describe us? It describes the problem we exist for. Olivares AI’s first job is to make sprawl visible: it discovers the agents, models, MCP servers and tools in your estate and builds a read/write access map of what each can reach — read-first, minimal-data, on your infrastructure. The Permitted-vs-Observed diff then turns “we have a lot of agents” into “here are the ones using access nobody granted.” Sprawl is the disease; an accurate, attributed inventory is the first treatment.

Guardian agents

What it means. Gartner’s term for AI capabilities that monitor, oversee or intervene on other AI agents. Gartner projects guardian-agent technologies will account for 10–15% of the agentic AI market by 2030 (Gartner press release, 2025; see sources).

Does it describe us? Carefully. Olivares AI delivers the governance and oversight outcome the category is about — observing agent behaviour, diffing permitted against observed, gating actions deny-closed, and recording everything in a tamper-evident ledger. But we are not an autonomous runtime agent that reasons about other agents in the request path. We are a read-first control plane that sits outside the data path: we observe through telemetry, native audit logs and an eBPF kernel backstop, and we enforce at well-defined gates (approvals, the Claude Code hooks PEP, kill switches) — not by inserting an AI proxy into every call. If “guardian agent” means supervisory governance over your agent estate, yes. If it means an LLM standing guard inline, that is a different architecture, and we will not claim it.

AI TRiSM

What it means. AI TRiSM — AI Trust, Risk and Security Management — is a framework coined and owned by Gartner for managing the trust, risk and security of AI across its lifecycle. As commonly summarised, it spans governance and runtime inspection & enforcement of AI, alongside information governance and infrastructure security.

Attribution note

The AI TRiSM framework, its layer taxonomy and any definitions are Gartner proprietary research. Public restatements (including layer names and diagrams) typically originate from licensed reprints. We describe AI TRiSM at the theme level and map our capabilities to those themes; we do not reproduce Gartner’s exact model, claim conformance to it, or imply a Gartner endorsement.

How we map to it (theme level).

• Governance — policy authoring, risk classification (EU tier × NIST function), approvals/HITL, manage-as-code, and the compliance module’s framework catalog.
• Runtime inspection — the access map and Permitted-vs-Observed drift, guardrail/anomaly findings, session timelines — all read-first and out-of-band.
• Runtime enforcement — deny-closed gates where we do sit in a decision path: approvals, the Claude Code hooks PEP, MCP tool gating, kill switches.
• Information governance — PII/sensitivity discovery over governed knowledge bases, data-residency attestation, retention and legal-hold.

We use AI TRiSM as a map of the problem space a buyer already knows, to show coverage — not as a badge.

Discover / observe / govern / secure

What it means. The verb sequence analysts and vendors use to describe the AI- governance lifecycle: first discover what exists, then observe what it does, then govern what it is allowed to do, then secure the whole estate.

Does it describe us? Yes — it is close to our own product narrative, which is worth stating in our exact terms so the mapping is honest:

Analyst verb	What Olivares AI actually does
Discover	Inventory of agents, models, MCP servers and tools across the estate.
Observe	The R/RW access map — read-first, minimal-data, with per-edge attribution confidence; cooperative paths (OTel, hooks) corroborated by native audit (pgAudit, CloudTrail) and an eBPF backstop.
Govern	Permitted-vs-Observed drift, policy + approvals/HITL, deny-closed actuation gates, manage-as-code.
Secure	Guardrails, the tamper-evident audit ledger, kill switches, compliance evidence — and the posture that the whole thing runs self-hosted, with no mandatory telemetry and no control-plane egress by default: what crosses your perimeter is what you configure to cross it (model APIs, the SIEM/webhook outputs you wire, an external embedding provider if you provision one).

The honest caveat that runs through all four: fidelity is tiered. Observation is clean for SQL databases, object stores and warehouses; lossy for document and vector stores; and not achievable passively for some systems. The map shows its confidence rather than inventing attribution it does not have.

The three lanes this vocabulary points at

Strip the labels away and the same three differentiators remain — the lanes the market has left open and the copy should keep returning to:

1. Ground truth from the data plane. We do not take an agent’s word for what it touched. We correlate the cooperative signal (OTel, MCP, hooks) against the system’s own ledger — pgAudit classifying reads vs writes, CloudTrail exposing object-store access — and an eBPF kernel backstop for the non-cooperative case. That correlation is what makes Permitted-vs-Observed a fact, not a self-report.
2. Deny-closed enforcement on the local dev agent. Most tools only observe Claude Code. Olivares AI also governs it: the hooks PEP turns policy into a deny-closed decision at the agent, not an after-the-fact log line.
3. Sovereignty. Self-hosted, source-available AGPL — the data plane never leaves your boundary and there is no SaaS control plane in your compliance path.

Every term above is in service of those three. When a page here uses an analyst word, it is to meet the buyer where they are — then to point back at one of these three things the product genuinely does.