Agentic Coding: Basics and Best Practices

Genesis Mission Platform — Foundations session F5 · October 15, 2026

Build with tests, make it fast, give it a face.

The one idea

You set the goal and the direction. The agent executes. You verify. The agent makes code cheap to write; reviewing, maintaining and trusting it are not. Use the agent to raise the bar — better tests, better performance, a better interface — not the line count.

What you will be able to do

1. Run a coding agent (Claude Code, Codex CLI, OpenCode, Cursor or Kilo Code) on a real repository.
2. Steer it with four layers: instructions (AGENTS.md or CLAUDE.md), skills (SKILL.md), memory and permissions.
3. Let tests be the stopping rule, so "done" means the tests pass, not that the agent says so.
4. Go from correct, to fast, to usable, see hands-on demos of the session.

Where to go next

• Getting started: pick one tool and begin.
• Shell setup: switch Claude Code between MAG and ALCF from bash or zsh, on any Linux machine or supercomputer.
• Run in a container: every tool and editor in one image, with your project and saved settings mounted.
• Best practices: the habits and the evidence behind them.
• Tools cheat sheet: the same move in each tool.
• Working with multiple agents: divide work, run sessions in parallel, integrate.
• Skills: procedures the agent follows when a task calls for them; where to find good ones, including Genesis Skills.
• MCP servers: connect the agent to tools and data, such as current docs, a browser or GitHub.
• Hooks: commands that always run at set points, such as formatting after every edit.
• Models and budget: how to choose, and how to see what you spend.
• Custom endpoints: argo-proxy, argo-shim and ALCF Inference Endpoints.
• Hands-on demo: demo repositories.
• Helpers: who to ask, and about what.

Model access for Genesis Mission teams: AmSC Model Access Gateway (MAG).