AI Agents, Demystified
Seven words you keep hearing — chat history, context, skill, tool call, agent, sandbox, MCP — and the plain architecture behind each one.
There is no magic here. Every one of those seven words — chat history, context, skill, tool call, agent, sandbox, MCP — is a boring engineering solution to a boring engineering constraint, and they build on each other in exactly that order:
- The model is stateless, so we invented chat history.
- History grows without limit, so we care about context.
- Context is finite but instructions are many, so we load them on demand as skills.
- The model can only emit text, so we invented tool calls.
- One tool call is rarely enough, so we wrapped it in a loop and called it an agent.
- The loop now runs code we never reviewed, so we need a sandbox.
- Everyone writes the same tools over and over, so we standardised on MCP.
This talk walks through each concept in that order, with the architecture behind it, so none of these words have to be magic anymore.