The coincidence is not an accident: a model's context is finite, and answer quality drops noticeably once reading code, planning, building and reviewing are all dumped into one window at the same time. Splitting the work into narrow subtasks, handing each to a separate worker in its own window, and taking back a compressed output rather than a raw log is a solution that engineers at seven different companies arrived at independently of one another. That is the right way to read this reference: not as seven separate inventions, but as seven implementations of the same engineering necessity.

The first table folds these vocabularies into one shared coordinate system: what the orchestration is called in a given environment, how delegation actually happens, and how context discipline is kept. It reads more naturally by column than by row. The delegation column shows exactly how the lead agent hands off work - through a separate subagent, through a handoff that fully transfers the lead, or through an ordinary tool call. The context-discipline column shows what comes back up once a worker has finished - a compact summary, not a raw log of everything it read and did.

The second table is factual, and it is where the main split between environments becomes visible: environment type, exactly where the config lives, where subagents are declared, and which model and architecture stand behind the environment. Some ship their own CLI shell with their own config and their own subagents - Claude Code, Codex, Gemini, Kimi, Qwen, and GLM with its own ZCode. Others have no shell of their own at all: the model is wired as a backend to someone else's host, most often the same Claude Code, and all orchestration is taken from there. GLM can work both ways; DeepSeek only the second.

From here the reference proceeds through the environments in the same order: first the ones with their own shell, config and subagents, each with its own move, where the setting lives, which model stands behind it, and what to do with it in practice. Then the environments that are wired into Claude Code as a model backend, where the orchestration is no longer their own but the host's. It is worth keeping both tables in view on a first read: they give a map to orient by, without having to hold all seven environments in your head at once.