---
title: "NVIDIA NemoClaw for Agent Optimization"
publisher: "nvidia"
type: "blueprint"
updated: "2026-10-20T07:10:06.824Z"
description: "Find agent failures in agent traces, build task-specific evaluations, and validate harness changes before shipping them."
canonical: "https://build.nvidia.com/nvidia/nemoclaw-for-agent-optimization"
---

## What is NemoClaw for Agent Optimization?

This is an open blueprint that shows how to improve an agent’s accuracy and cost profile by iterating on its harness — the prompts, tool descriptions, and routing logic the model depends on but never sees during training.

The blueprint  uses NVIDIA NeMo Compass to surface failure patterns in recorded agent runs, and NVIDIA NeMo Eval Author to build Harbor evaluations against those specific failures. You make a targeted change to one part of the harness, run both arms, and compare results before committing the change.

Two runnable examples are provided—one using a Hermes enterprise assistant agent, and one using a LangChain agent applied to FreeCAD CAE work—each with saved traces, runnable evaluation tasks, and a saved baseline/candidate comparison you can inspect without a live model run. An ETHOS.md file defines the agent’s intended behavior and is shared across both trace analysis and evaluation authoring, so the two phases stay aligned.

## Why use it?

- Root-cause from traces, not guesses. NeMo Compass examines recorded runs and surfaces where the agent called the wrong tool, misrouted a request, or produced output that violated its intended behavior.
- Evaluations tied to real failures. NeMo Eval Author builds Harbor tasks directly from the issues NeMo Compass finds, so your evals cover the actual work the agent does—not synthetic scenarios.
- Guard against overfitting. The blueprint uses a development/held-out split and validates that fixes pass repeatedly before accepting them.
- A/B comparison built in. The comparator validates run identity, model, timeout budgets, and task checksums across all jobs before accepting a comparison, so you’re measuring the harness change, not run variance.
- Start from a working example. Two examples are included: a Hermes personal assistant agent and a LangChain agent for FreeCAD CAE work.  Swap in your own traces, tools, and ETHOS.md to adapt either one.

## Example workflow

![Agent optimization workflow: Hermes Agent runs in OpenShell, NeMo Relay captures traces for NeMo Compass and NeMo Eval Author, and findings and evaluations guide harness optimizations.](https://assets.ngc.nvidia.com/products/api-catalog/nemoclaw/nemoclaw-for-agent-optimization/workflow.png)

1. Run one of the included examples in OpenShell and capture baseline traces. The personal assistant example uses Hermes Agent; the FreeCAD CAE example uses LangChain. The process applies to any open harness.
2. Open NeMo Compass and identify a failure pattern—a wrong tool call, a misrouted request, or a violated behavior defined in ETHOS.md.
3. Use NeMo Eval Author to write a Harbor task that tests for that specific failure.
4. Change one part of the harness (a system prompt, a tool description, a routing rule).
5. Run both arms and use the comparator to validate the A/B result.
6. Keep the change if it passes on the held-out set; otherwise revise and repeat.
7. Apply the same loop to your own agent by replacing the example traces, tools, and ETHOS.md.