GitHub Agentic AI Developer Course Notes

A cloud agent is used when a repository task needs more than a chat answer: Copilot receives repository context, works in an isolated branch, commits proposed changes, and opens a pull request for review. GH-600 questions often separate this workflow from ordinary chat because the cloud agent produces repository artifacts rather than only suggesting code inline.

Tool interaction begins with deciding which external capabilities the agent may call. MCP connects Copilot to tools and data sources through servers, while Copilot SDK custom agents define which tools are available and what permissions apply during execution.

Short-term memory is the context available inside the current agent session or conversation. Use it for details that matter only while the current task is active, such as the files already inspected, a temporary decision, or the latest reviewer instruction.

Evaluation starts by naming what success looks like before accepting agent output. Quantitative signals include test results, lint counts, changed files, and detected regressions; qualitative signals include code clarity, maintainability, security risk, and whether the pull request actually solves the requested problem.

Multi-agent coordination uses runtime delegation when one agent cannot sensibly handle every part of a task. A coordinator or SDK runtime chooses the agent profile that best matches the requested work, tool needs, repository context, and current step.

Guardrails define what an agent may do without approval and what must stop for human judgment. GH-600 questions usually tie guardrails to repository patterns, autonomy levels, branch rules, and review requirements rather than to vague reminders about being careful.