Module 7 Lab — Agent Harnesses & Skills

Why it matters

Every tool and skill in a prompt costs context on every call. Measuring that cost turns lazy loading from a preference into a design decision.

What to practise

Build a harness around Nemotron from a short system prompt, four tools — read_file, write_file, edit_file, run_bash — and a loop. Estimate the prompt-and-tool overhead of two configurations with tiktoken — a comparison proxy, not the model’s billed count — then implement lazy loading. Write a dataset-profiler SKILL.md whose description triggers loading, install and verify the signature of NVIDIA’s accelerated-computing-cudf skill, and let the agent write a new skill from its own transcript.

How it fits

The four tools behind one loop are 1.8.1’s separation of model, tools and orchestration. The context tax and lazy loading are the deferred tool and skill loading in 1.8.3. A run_bash tool is also where 1.8.2’s sandboxing and deadlines matter most.