HOW AI AGENTS WORK — PRIMARY SOURCES
Prepared 24 October 2026. The visuals explain durable concepts; named-product availability and capabilities can change.

Agent and harness architecture
OpenAI — Codex as a platform
https://developers.openai.com/blog/codex-as-a-platform

Skills, extensions, and platform-specific features
Anthropic — Claude Code features overview
https://code.claude.com/docs/en/features-overview

MCP architecture and tool/data connections
Model Context Protocol — Architecture overview
https://modelcontextprotocol.io/docs/learn/architecture

Tokens, context, and retrieval-augmented generation
Anthropic — Claude glossary
https://platform.claude.com/docs/en/about-claude/glossary

Inference
Google Cloud — What is AI inference?
https://cloud.google.com/discover/what-is-ai-inference

Usage, observability, and cost
OpenAI — Agents API observability
https://developers.openai.com/api/docs/guides/agents-api/observability

Browser extension work
OpenAI — ChatGPT Chrome extension
https://learn.chatgpt.com/docs/chrome-extension

Agent delegation and handoffs
OpenAI Agents SDK — Orchestrating multiple agents
https://openai.github.io/openai-agents-python/multi_agent/

Editorial notes
A permission check uses authorization already given; it does not imply asking for approval before every step.
A model may finish or ask a person without using a tool.
Retrieval and helper agents can improve a workflow but do not guarantee accurate output.
Plugins may bundle skills, tools, and connections; exact contents vary by platform.
Neither MCP, RAG, plugins, nor multiple agents is required for every workflow.
Income examples describe service possibilities, not income promises.
