Inside AI agents: build a spy for GitHub Copilot, Cursor and Claude
An agent's transcript doesn't tell the whole story. We build its flight recorder live.
Code-generating AI agents can read source files, invoke MCP server tools, run commands and delegate tasks to sub-agents. But the transcript of their conversations doesn't reveal the whole story.
Did the agent load the expected skill? Which MCP tool did it call, and with what arguments? Was that call part of the main conversation or the work of a sub-agent running in parallel? How long did each step take, and at what cost?
GitHub Copilot, Cursor and Claude Code offer official hooks covering sessions, prompts, tools, files, MCP calls, context compaction and sub-agents. In this demo-driven session, we'll use them to build a passive flight recorder in .NET.
You'll then see how to correlate session, turn, tool and sub-agent identifiers to reconstruct the agent's activity and summarise its use of tools, MCP and skills.
No traffic interception, no undocumented APIs and no vendor marketing: only officially supported hooks, working code and an honest look at the limits of observability.