The leading development assistants use different command names, different configuration files and different models under the hood, but the load-bearing move is the same across all of them. The lead agent does not write code itself - it holds the whole task, from the initial statement to the finished result, and hands narrow slices of it to subagents. Each subagent works in its own clean context, sees neither the orchestrator's history nor a neighbor's subtask, does exactly one thing, and returns not the raw stream of everything it read and did but a compact summary. The orchestrator collects these summaries one after another and moves the task forward on their basis.

The move rests on just three rules, and together it is convenient to call them context discipline. A separate window for each subagent protects the model's attention: it does not drown in someone else's logs and does not confuse its own narrow subtask with a neighboring one that someone else is solving at the same moment. A summary instead of a full dump saves tokens at every step and, more important than the savings themselves, leaves the orchestrator room to think once the workers have returned, rather than while parsing their output. Stop conditions keep loops from spinning in place when a fix does not move the result. The difference between a governed multi-agent system and one overloaded conversation comes down to exactly these three rules, not to how many models are involved.

The reference is laid out environment by environment, and each one describes the same set of things: which pattern stands behind its orchestration, where the config lives, where subagents are declared, which model and architecture stand behind the environment, official links to the primary sources, and a ready "configure yourself" meta-prompt. The meta-prompt is simple by design: you paste it as the first message at the project root, and the agent reads the listed documentation itself, then writes its own config and its own subagents for the specific project, showing the plan before anything is written to disk. The meta-prompt is a starting frame, not dogma: it is worth adjusting to the particulars of your own project rather than executing word for word.

Environments in this reference fall into two groups, and the line between them runs along product architecture, not quality. Some have their own CLI shell with their own config and their own subagents - there the pattern is assembled with the environment's native tools. 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, through an Anthropic-compatible API, and in that case orchestration is taken entirely from the host, with nothing configured anew.