Architecture: An Agent Optimization Loop

The NemoClaw agent optimization blueprint gives you a reproducible loop for finding and fixing harness inefficiencies—the prompts, tool descriptions, and routing logic around the model—without modifying the model itself. . The example agent runs a task, records what happened, and produces a baseline. Use that evidence to pinpoint a failure, write a test for it, and measure whether a change helps.

Core architecture

The blueprint has three groups of components: one to run the agent, one to capture what it does, and one to analyze and test the changes.

ETHOS.md is a plain-text file you write that defines the agent's intended behavior. Both NeMo Compass and NeMo Eval Author read from it, keeping analysis and evaluation aligned on what “correct” means.

How the blueprint works

The optimization target is the harness—the system instructions, tool descriptions, and routing logic the model depends on but never saw during training. The loop is: record what the agent did, find what went wrong, write a test for it, change one thing, and measure the result.

Two examples are included: a Hermes personal assistant agent and a LangChain agent for FreeCAD CAE work. Both ship with recorded traces, runnable evaluation tasks with proof receipts, and a saved baseline/candidate comparison you can inspect without running a model. A shared ETHOS.md defines the agent’s intended behavior and is referenced by both the trace analysis and evaluation steps.