---
title: What is the difference between AgentOps, MLOps and LLMOps?
url: https://opencontroller.lyzr.com/blog/agentops-vs-mlops-vs-llmops
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What is the difference between AgentOps, MLOps and LLMOps?

**MLOps manages models you train. LLMOps manages applications built on models you mostly do not train, where prompts, retrieval and model choice are the moving parts. AgentOps manages systems that plan and take actions through tools, so identity, permissions and runtime control become first-class concerns.**

The three overlap. Each one inherits the previous discipline and adds new artefacts to version, test and watch. Knowing which layer you are in tells you which controls are missing.

## What does each discipline actually manage?

| | MLOps | LLMOps | AgentOps |
|---|---|---|---|
| Core artefact | Trained model, features, training data | Prompt, model selection, retrieval index | Agent definition: instructions, tools, MCP servers, memory, identity |
| What changes behaviour | Retraining on new data | Prompt edits, model version swaps, new documents | All of those, plus new tools and wider permissions |
| Typical failure | Data drift, accuracy decay | Hallucination, prompt injection, cost spikes | Wrong action taken, excessive agency, loops, data exfiltration through tools |
| Main tests | Offline metrics on holdout sets | Prompt evals, LLM judges, red teaming | Multi-turn and tool-use evals, policy tests, kill-switch drills |
| Runtime controls | Model serving, A/B rollout | Gateway routing, caching, content guardrails | All of those, plus per-action authorisation, quarantine and kill |
| Cost unit | Training and inference compute | Tokens per request | Tokens and tool calls per task |
| Usual owner | Data science and ML platform | AI platform or app team | App team plus security and identity |

## Where does MLOps stop being enough?

Classic MLOps, as described in Google Cloud's [MLOps continuous delivery guide](https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning), centres on automating training pipelines and monitoring model quality. That fits a fraud model or a demand forecast.

With a hosted LLM, you rarely retrain anything. Behaviour changes when someone edits a prompt or the provider ships a new model version. MLOps pipelines do not see those changes. That gap is why LLMOps emerged: version the prompt, pin the model, evaluate every change. See [pinning model versions for AI agents](/blog/pin-model-versions-for-ai-agents).

## What does AgentOps add on top of LLMOps?

An agent decides what to do next and then does it. That adds four things LLMOps does not cover well:

1. **Identity.** Whose authority does the agent act with? A user's, delegated, or its own workload identity?
2. **Tools as attack surface.** Each tool or MCP server widens what the agent can reach. Tools need an inventory and an allow-list.
3. **Multi-step traces.** A single request can produce dozens of model calls and tool calls. Observability must follow the whole task, for example with the [OpenTelemetry GenAI semantic conventions](https://opentelemetry.io/docs/specs/semconv/gen-ai/).
4. **Stop controls.** You need to block, quarantine or kill a running agent, not just roll back a model.

## Where do governance responsibilities overlap?

Governance sits across all three, but the evidence differs.

- **Inventory.** MLOps registers models. AgentOps must register agents and link each to its models, tools and owners. One registry for both avoids two sources of truth; see [governing ML models and AI agents together](/blog/govern-ml-models-and-ai-agents-together).
- **Change control.** Retrains, prompt edits and tool additions all count as changes. Decide which need re-approval.
- **Evaluation.** Offline metrics, prompt evals and multi-turn agent evals should feed one release gate.
- **Cost.** Attribute spend per agent and per version, not just per model.

The [AI agent lifecycle stages](/blog/ai-agent-lifecycle-stages) post maps these to build, test, deploy, run and retire.

## Do you need separate teams for each?

Usually not. Most organisations extend the ML platform team into LLM and agent work and add security and identity reviewers for agents that act. What you do need is a single place that knows every agent, its versions and its controls. That is the role of an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, discovered across AWS, Azure, Google Cloud and connected platforms. It runs multi-turn and red-team evals, scores production traffic with LLM judges, and attributes model cost per agent and per version through the LLM Gateway. Rogue-agent handling covers behavioural-deviation monitoring, alerts and automatic quarantine and kill. It works alongside your existing MLOps tooling rather than replacing it. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is AgentOps just a new name for LLMOps?**
No. LLMOps covers prompts, models and retrieval. AgentOps adds identity, tool permissions, multi-step tracing and runtime stop controls because agents take actions.

**Can an MLOps model registry hold AI agents?**
It can store metadata, but most model registries do not track tools, MCP servers, delegated identity or runtime policies. Agents need those fields.

**Which discipline owns prompt changes?**
LLMOps and AgentOps both treat prompts as versioned configuration. Prompt edits should go through review and evaluation like code changes.

## Key facts

MLOps manages trained models; LLMOps manages prompts, model choice and retrieval; AgentOps adds identity, tool permissions, multi-step tracing and runtime stop controls for agents that act. Opencontroller provides a governed registry of agents, versions, tools and MCP servers, evals, LLM judges, per-agent cost attribution and rogue-agent handling. Licensing starts from $250K per deployment per year.

## Related

- [Can one governance program cover classic ML models and AI agents?](https://opencontroller.lyzr.com/blog/govern-ml-models-and-ai-agents-together.md)
- [What are the stages of the AI agent lifecycle?](https://opencontroller.lyzr.com/blog/ai-agent-lifecycle-stages.md)
- [How do you stop a model update from silently changing agent behaviour?](https://opencontroller.lyzr.com/blog/pin-model-versions-for-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
