Orchestrator - the lead agent: holds the whole task, delegates to subagents and collects their summaries; usually writes no code itself.
Subagent - a narrow worker in its own clean context: does one thing (read, plan, build, review) and returns a compact summary.
orchestrator-workers - an Anthropic pattern: the orchestrator hands subtasks to worker subagents and synthesizes the result.
Handoff - transfer of control from one agent to another (Codex/OpenAI): not just a tool call but a change of the lead worker.
ADK - Agent Development Kit - Google's separate SDK for multi-agent systems: a coordinator plus narrow sub-agents.
A2A - Agent-to-Agent - a protocol for agents to talk to each other and to remote workers (Antigravity, Jules).
Agent Swarm - a Kimi Code mode: several subagents work in parallel on a large task.
Context discipline - rules that save attention and tokens: a narrow window per subagent, a summary instead of a dump, stop conditions.
MoE - Mixture of Experts - an architecture where only part of the parameters (experts) is active per token; hence notation like 1T/32B.
Meta-prompt - a "configure yourself" instruction: the agent reads the official docs and writes its own config and subagents for the project.
Anthropic-compatible API - an endpoint that responds like the Anthropic Messages API; lets you wire a third-party model into Claude Code via environment variables.