The life agentic

Getting started with agentic LLM use

This section collects practical information for students who want to get started with agentic use of LLMs in a way that fits this course and the surrounding tooling.

If you want to see what these tools actually do before deciding whether the setup is worth it, see what an agentic session looks like first. The rest of the guide takes the pieces of that session apart.

Several common ways to begin are:

  1. If GitHub student benefits are available to you, start there and use GitHub tools such as Copilot.
  2. If you want a paid fallback, start with Claude Code and a Claude Pro subscription.
  3. Move to command-line workflows such as GitHub Copilot CLI (copilot) and opencode for repository work.
  4. Use the Python package llm when you want shell pipelines, scripts, or Python integration.

As you read the three main coding-agent guides, notice that they keep returning to the same problems: planning before editing, managing a limited context window, isolating side work, and separating instructions from memory. The page Agentic concepts reconnects those examples after you have seen the tools themselves, and also explains why this course puts so much weight on local CLI agents rather than only browser chat. Many students have already tried tools such as ChatGPT in the browser. That is a fine starting point, but work becomes much more efficient when an agent can inspect the actual files, search the project, run commands, and use tools directly. That matters not only for editing code, but also for repository-grounded questions, planning, and implementation ideas.

Intended learning outcomes

Across this section, students should be able to:

  1. Distinguish between an LLM tool, an underlying model, and a model-access route.
  2. Choose an appropriate access route (for example GitHub, Claude, direct API, Zen, or Ollama) based on availability, cost, and desired workflow.
  3. Choose an appropriate tool for a task by distinguishing between coding-agent workflows and scripting or pipeline workflows.
  4. Explain why repository-grounded terminal agents support different kinds of work than ordinary browser chat.
  5. Use a coding agent in a way that separates planning from editing.
  6. Recognise context-window limits and use strategies such as fresh sessions or compaction when context quality degrades.
  7. Explain how subagents or separate runs can isolate side work and keep a main task focused.
  8. Distinguish between persistent instructions and reusable memory or context.
  9. Use AGENTS.md and skills appropriately as reusable instruction layers in a project workflow.
  10. Explain when local models via Ollama are a good fit and when hosted tools are the better choice.
  11. Use llm for a repeatable shell or Python-based workflow.
  12. Critically verify model output against files, tools, and project context before relying on it.
  13. Choose an appropriate model family for a task by weighing quality, speed, cost, and local-versus-hosted constraints.
  14. Identify the main legal, privacy, and reliability risks of using LLM tools on real work, and choose mitigations such as local models, verification, or restricted prompts.
  15. Explain how copyright, licence contamination, and attribution interact with code generated by an LLM.
  16. Reflect critically on the wider consequences of agentic LLM use — on skill, comprehension, ownership, and the social context of software work — and engage informedly with the public discussion around these consequences.

Start here

Common starting points

GitHub is still worth trying when student benefits are available, because it can be the easiest low-cost path to Copilot tools.

Claude Code with Claude Pro is the clearest paid alternative when you want to start immediately with one polished coding agent.

If you are unsure how tools, subscriptions, APIs, and local models fit together, start with Model access.

If you specifically want a local or offline route, read Local models and Ollama after that.

Skills and AGENTS.md become useful soon after that, because they let you keep reusable instructions in files instead of rewriting them in every chat.

Guides in this section

These pages cover:

This site is open source. Improve this page.