What an Agent Management Platform Actually Is (and Is Not)

An agent management platform is one console to register, monitor, govern and optimise every AI agent a business runs, whichever vendor or harness each agent lives in. That is the whole definition. It is a control plane, not another agent, and its defining property is neutrality: it has to manage the agents you bought, not just the ones somebody built.

Gartner formalised the category in 2026, which means every vendor with a dashboard is now claiming the label. So this post is the plain version: what the category actually is, what it is not, and how to tell the difference before you buy one.

Why the category exists: agent sprawl

Ask a mid-sized business how many AI agents it runs and nobody can answer. Marketing has a ChatGPT workflow. Engineering runs Claude agents. Sales bought something with an AI SDR bolted on. The security stack shipped its own copilot last quarter, whether anyone asked for it or not. Each one has its own console, its own bill, its own idea of logging, and its own blast radius.

This is agent sprawl, and it is the new shadow IT. The difference from the SaaS sprawl of the 2010s is that these systems do not just hold data, they act: they send emails, update records, move money through approval chains. An unlisted SaaS tool is a compliance gap. An unlisted agent is an unsupervised employee.

The register comes first. You cannot govern what you cannot list.

What an agent management platform does

Four functions, in order, because each one depends on the one before it.

**Register.** Every agent in the business gets an entry: what it is, who owns it, what it is allowed to touch. Including the embedded ones that came inside SaaS products. The register is the difference between an estate and a rumour.

**Monitor.** Every AI conversation measured on one set of axes across every vendor: tokens in and out, tools called, latency, cost per interaction. Not five dashboards with five definitions of a request. One.

**Govern.** The consequential stuff. Actions that change the business, a payment, an outbound email, a record update, wait in a human approval queue before they run. Boundaries are enforced at the tool call. An audit trail records who approved what. And any agent can be paused or stopped instantly. No agent acts alone, and every agent is killable.

**Optimise.** Once everything is measured the same way, the levers become obvious: route each request to the model that handles it best instead of sending everything to the most expensive one, restructure the retrieval that is stuffing whole histories into every prompt, parallelise the tool calls that run serially. The spend drops because the estate is finally legible.

What it is not

Three neighbouring categories get sold under this label. All three are real tools. None of them is the thing.

**It is not observability.** Langfuse, LangSmith, Arize and Datadog instrument the agents your engineers build. They are excellent at traces and evals, and blind to any agent nobody instrumented, which in a real business is most of them. They also barely touch approvals: watching an agent is not governing it.

**It is not a gateway.** Portkey, Kong, LiteLLM and TrueFoundry govern traffic that routes through them. Useful exactly there, and structurally unable to see an agent embedded in a SaaS product you cannot repoint. They govern the wire, not the worker.

**It is not a platform control tower.** Microsoft, Salesforce and ServiceNow each ship genuinely deep admin tooling for their own agents. The operative words are their own. A control plane that only covers one vendor's estate is a very good dashboard with a category name attached.

The test is one question: can it register, monitor and govern an agent it did not build, on a platform it does not own? If the answer is no, it is one of the three above, whatever the label says.

How agents actually connect

The honest answer is: at different depths, depending on how much each agent exposes. We wrote the full model in the [AIOS whitepaper](/aios-whitepaper), but the short version is four tiers. Agents you build or self-host route inline through a gateway, full fidelity, budgets enforced before spend happens. Cloud runtimes like AWS AgentCore stream native telemetry, near-full fidelity with no rebuild. Closed platforms are polled through the vendor's own admin and audit APIs. And true black boxes, the embedded copilots that export nothing, are still registered, owned, bounded and logged.

The point of the tiers is that nothing in the estate is left ungoverned, and nothing gets rebuilt to make that true. Any vendor who tells you every agent must be rebuilt onto their framework is selling you an agent builder with a governance tab.

Who actually needs one

A rough threshold: the day the business runs more than five agents across more than one vendor, or the day any agent can take an action a human would be fired for taking unsupervised. For most mid-market companies in Australia and New Zealand, at least one of those days already happened, quietly, without a decision being made.

We build one. It is called the [Sentry Agent Management Platform](/agent-management-platform), it is the neutral kind, and it is delivered as a managed service with a named team rather than software you have to staff. It is currently in free beta with no lock-in, which is also our answer to the how-do-we-know-it-works question: connect an agent and see.

Where to start

Not with the platform, honestly. Start with the register, and the register starts with knowing what your business actually runs. Our [AI Opportunity Audit](/ai-opportunity-audit) maps your tools, teams and data into one graph in about two minutes, free, in your browser. It will not list your agents for you, but it makes the sprawl visible, and the follow-up conversation is where the estate gets counted.

You cannot govern what you cannot see. Register everything, then decide.