Building Coding Agents

Coding agents — Cursor, Claude Code, OpenCode and their kin — are the same tool-calling loop as any agent with a filesystem tool set. The AISL material builds them from scratch across three workshops that escalate: a project bootstrapper, a skills layer, and the combined current version.

The anatomy: tools, template, prompt

Building a Coding Agent with Python and Django builds a small project bootstrapper: it copies a working Django template, then reads and writes files through tools until the generated app runs. Three ideas carry the whole workshop. Tools are ordinary Python functions the model calls. Templates beat empty directories — Lovable and friends start from prepared projects too. And behavior lives in the developer prompt, which deserves most of your attention — Prompt Engineering collects how the material treats prompts as an engineering surface.

Skills and slash commands

Skills and Slash Commands from Scratch adds the reusable behavior layer. A skill is a folder whose SKILL.md the agent loads when a request matches its description; a slash command is one you invoke explicitly, rendered into the prompt before the agent acts. The workshop parses and loads skills in Python, exposes them as one skill tool, and mirrors how Claude Code and OpenCode implement the same ideas.

The current combined version

Coding Agent with Skills, from Scratch combines and updates both: tool calls with the Responses API, a five-tool agent — read, write, tree, bash, search — inside a Django project template, the skills system, and a port to PydanticAI. If you take one workshop on this topic, take this one. It was delivered at MLCon 2026 Amsterdam.

Where the agents live

Long agent sessions need a home that survives disconnects: Setting Up a Remote Environment for Agentic Workloads prepares an Ubuntu box — SSH hardening, runtimes, closed ports — and keeps Codex, Claude Code, OpenCode, and Grok Build in persistent tmux sessions you can reach from any device. Provisioning, session isolation, autonomy boundaries, and device access get their own page: Agentic Development Environments.

Using them well

Building agents is half the story; driving them is the other: see AI Coding Tools for the tool landscape and the workflow experiments — including slash-command loops in Experiments with Claude Code.

Where to go next

• The loop underneath: Agents
• Compare the ready-made tools first: AI Coding Tools
• Evaluation and Monitoring — the coding-agent workshop names evaluation as the deliberate gap it leaves you