About Agents

An agent on NeMo Platform is a workflow that calls tools, talks to models through the local platform, and runs as a managed service you can deploy, invoke, evaluate, and optimize as a unit. Agents are defined as NeMo Agent Toolkit (NAT) workflows and are managed through the nemo agents command group.

About NeMo Agent Toolkit

NVIDIA NeMo Agent Toolkit is a flexible, lightweight, and unifying library that allows you to easily connect existing enterprise agents to data sources and tools across any framework.

NeMo Platform uses NAT as the runtime wrapper around your agent so the platform can deploy it, evaluate it, optimize it, and route its model traffic through shared infrastructure. For the toolkit itself, see the NeMo Agent Toolkit documentation.

Agent Definition

An agent’s behavior is described by a NAT workflow YAML with three top-level sections:

ReAct is a common agent pattern where the model alternates between a reasoning step and a tool call until it has enough information to answer. It is a good default when you want the agent to decide which tool to use next based on what it has already learned. A minimal ReAct agent looks like this:

1
functions:
2
  wiki:
3
    _type: wiki_search
4
  clock:
5
    _type: current_datetime
6
7
llms:
8
  llm:
9
    _type: openai
10
    model_name: nvidia-nemotron-3-nano-30b-a3b
11
12
workflow:
13
  _type: react_agent
14
  tool_names: [wiki, clock]
15
  llm_name: llm

The Inference Gateway is the local platform’s model proxy. Agents send model requests to it instead of to provider APIs directly, so the platform can resolve model names, attach credentials, and route through middleware on the agent’s behalf. Two conventions apply when a config targets a deployed agent:

• Model names use the Inference Gateway entity form, with slashes and dots converted to hyphens (nvidia/nemotron-3-nano-30b-a3b becomes nvidia-nemotron-3-nano-30b-a3b). The Inference Gateway resolves the entity to the upstream provider that owns it.
• Leave base_url and api_key unset on openai and nim LLMs. NeMo Platform injects an Inference Gateway URL when it deploys the agent, and the gateway retrieves upstream credentials from the secrets service. Setting base_url explicitly bypasses both.

Agent Lifecycle

Agents are managed end-to-end through the nemo agents command group:

To run a workflow YAML directly without registering it on the platform, pass --agent-config <path> to nemo agents invoke or nemo agents run.

How It Works

A platform-managed agent consists of three components:

1. The agent entity. nemo agents create stores the workflow YAML as an entity in the workspace. The same configuration can be redeployed, evaluated, or optimized without re-registering it.
2. The deployment controller. nemo agents deploy passes the stored config to the Agents service controller, which launches a nat start fastapi process for it, assigns a port, watches its health, and tears it down on nemo agents undeploy.
3. The Agents gateway. Client requests reach the agent at /apis/agents/v2/workspaces/<workspace>/agents/<name>/-/<path>. The gateway resolves the agent to its current running deployment and proxies the request, including streaming responses. From a client’s perspective, the agent is an OpenAI-compatible endpoint owned by NeMo Platform.

Model traffic from inside the agent process routes back through the Inference Gateway, which resolves model entity names to upstream providers and supplies their credentials. This is why agent configs do not carry base_url or api_key values — the deployment injects the gateway URL automatically, and the gateway looks up the rest.

A virtual model is a platform-managed wrapper around one or more backend models. An agent can point at a virtual model entity while the virtual model handles routing, format translation, and guardrails behind the scenes — no changes to the agent’s workflow YAML required.

Applying Changes Through Candidate Agents

The Agents v2 API has no in-place patch for a stored agent config. When the optimize or secure workflows propose a change — a model swap, a routing split, a guardrailed virtual model — they apply it by creating a sibling agent with the new config, deploying it, and running an evaluation against it. The original agent stays untouched until you decide to promote the candidate.

The sibling-candidate pattern is the recommended way to apply changes: you always have a baseline to compare against, and you can roll back by undeploying the candidate.

Common Tasks

• Optimize Agents: analyze deployed agents for model routing, skill, prompt, and new-model opportunities.
• Secure Agents: check guardrail coverage and scan recent telemetry for sensitive data.
• Plugins and Skills: understand how agent, middleware, and coding-agent integrations extend the local platform.
• Agentic Metrics: evaluate tool use, goal completion, topic adherence, answer accuracy, and trajectories.
• Agent Configuration: use agents as online evaluation targets.