About Agents

An agent on NeMo Helix is an agent entity in a workspace. Once registered, the platform can deploy it, route requests to it, collect its traces, evaluate it, and optimize it. Agents run through NeMo Fabric, which gives every supported harness (DeepAgents, Claude Code, Codex, Hermes, Mini-SWE-Agent, NOOA, or a remote agent endpoint) the same configuration, lifecycle, and telemetry.

How It Works

Register. nemo agents register creates an agent identity. With --agent-config, it also stores the agent’s config and uploads the files beside a Fabric config to the agent’s <name>-ethos fileset.
Deploy. nemo agents deploy runs the agent as a local subprocess, a Docker container, or a Kubernetes Deployment.
Invoke. The Agents gateway routes each request to the agent’s active deployment and exposes it as an OpenAI-compatible endpoint, whatever the runtime.
Call models. Model traffic from the agent goes through the Inference Gateway, which resolves model and virtual model names and attaches provider credentials. Agent configs never carry provider API keys for platform-routed models.
Trace. Deployed agents export trajectories to Intake under the agent’s name by default, which feeds Insights, evaluation, and optimization.

Upgrade and adapter setup

For NeMo Helix 0.9, start with Fabric 0.5 Upgrade and Adapter Support: it covers the registration-command rename, host/task package compatibility, persisted Relay observability, gateway telemetry routing, and the difference between a shipped descriptor and an installed harness. Pi is an opt-in packaged integration, not an end-to-end release-gated harness in the 0.9 CI matrix.

For Mini-SWE-Agent 2.x, follow Run Mini-SWE-Agent for a complete install → register → deploy → invoke → session → cleanup journey.

Agent Lifecycle

Stage	Command
Register	`nemo agents register --name <name> [--agent-config <path>]`
Deploy	`nemo agents deploy --agent <name>`
Invoke	`nemo agents invoke --agent <name> --input "..."`
Chat	`nemo agents chat --agent-deployment <deployment>`
Run once as a job	`nemo agents execute --agent <name> --input "..."`
Analyze traces	`nemo compass analysis-runs create --agent <name>`
Optimize	`nemo agents optimize run-strategy --strategy <strategy> --agent <name>`
Update in place	`nemo agents redeploy --agent <name> --agent-config <path>`
Tear down	`nemo agents undeploy --agent <name>`, then `nemo agents delete <name>`

To try an agent.yaml without registering it, run nemo agents invoke --agent-config <path> --input "...".

Next Steps

Register Agents: from code, a container image, a remote endpoint, or traces.
Deploy Agents: choose a runtime, and build and publish agent images.
Execute Agents as Jobs: run an agent once, to completion.
Collect Agent Traces and Insights from Traces: find out what your agent does and where it fails.
Optimize Agents: generate, evaluate, and compare candidate agents.
Agent Governance: attack a sandboxed copy of an agent, then generate and validate guardrail and sandbox-policy fixes.