From Coding Assistance to Agentic Engineering

Title slide: GitHub Copilot, From coding assistance to agentic engineering.
The learning journey covers Assist, Ground, Delegate, Customize, Govern, and Specify.

Where are you today?

Audience poll ranging from never using GitHub Copilot to customizing agent tooling.
The goal is to move one step forward, using the least complex workflow that solves the task.

One product, three capability layers

Three GitHub Copilot capability layers: assistive, agentic, and agent customization.
Assistive experiences suggest and explain, agents plan and execute, and customization supplies context, tools, and controls.

Model, harness, tools, and controls

Diagram distinguishing the model, harness, tools, and controls in an agent loop.
The model reasons, the harness manages context and actions, tools perform work, and controls govern execution.

Which model? Start with Auto

Guidance to start with automatic model selection and choose manually only for a clear reason.
Manual selection is useful for known hard tasks, repeatability, policy, latency, context, or tool support.

Choose the surface by where work should happen

Comparison of GitHub Copilot surfaces for cursor, codebase, terminal, GitHub, and parallel workstreams.
Choose the surface based on where the work should happen, not which product appears most capable.

Better prompts are better task definitions

Weak and strong prompt comparison using goal, context, constraints, examples, and acceptance checks.
Better prompts are better task definitions grounded in files, tests, history, and tools.

Context shortcuts worth learning

Useful GitHub Copilot IDE and CLI context shortcuts and discovery commands.
Commands vary between surfaces; slash, at-sign, and help discovery are authoritative.

The customization ladder

Customization ladder from repository instructions to prompt files, skills, and custom agents.
Start with instructions and add complexity only when repeated work justifies it.

Built-in tools, MCP, and deterministic hooks

Comparison of built-in tools, Model Context Protocol tools, and deterministic hooks.
MCP extends external capability while hooks provide guaranteed lifecycle checks and controls.

Autonomy is a dial

Autonomy spectrum from exploration and planning to delegation and parallel isolated execution.
As autonomy rises, task clarity, tests, permissions, and review must rise with it.

GitHub is the control plane

GitHub control plane showing issue, branch, tests, pull request, human review, and protected merge gates.
Agentic work should pass through the same software delivery controls used for human contributors.

Agentic Engineering with Spec Kit

Spec Kit workflow from constitution and specification through plan, tasks, implementation, and convergence.
Durable specification artifacts coordinate work that outlives a single prompt while preserving quality gates.

Your homework

Closing slide with practical GitHub Copilot homework for new, regular, and advanced users.
Better context beats a more expensive model, greater autonomy needs stronger controls, and users remain accountable.