2.3.2. MCP, Agent Sessions, and Sub-Agents
💡 First Principle: MCP is how Copilot reaches outside its built-in knowledge to external tools and data; sub-agents are how a single task is split across parallel workers; agent sessions are the stateful container that holds the work together. One extends reach, one extends parallelism, one extends memory.
Model Context Protocol (MCP) is an open standard for connecting AI tools to external data sources and tools — databases, issue trackers, internal services, or custom servers. Copilot uses MCP to extend its context and capabilities; any compliant MCP server can act as a Copilot extension. Crucially, MCP replaced GitHub's proprietary Copilot Extensions (deprecated late 2025) and is GA across major editors. Enterprise admins govern MCP access through organization policy.
Agent sessions are the persistent, stateful runs in which agentic work happens (in the CLI or IDE). They carry context forward and, in recent updates, can run work in the background. Sub-agents let a primary agent delegate parts of a task to parallel workers for speed and for keeping each worker's context focused — the CLI's "fleet" behavior is an example.
💡 Key Point: If a question describes connecting Copilot to an external system (a database, a ticketing tool, a custom internal API), the mechanism is MCP. If it describes splitting one task across parallel workers, that's sub-agents.
⚠️ Exam Trap: MCP is an open standard, not a GitHub-proprietary, repo-only feature. Treating it as "GitHub's way to read your repo" misses the point — it connects Copilot to arbitrary external tools and data.
Reflection Question: A team wants Copilot to answer questions using data from their internal incident-tracking system. Which capability makes that possible, and who controls whether it's allowed in an enterprise?
