At Google orchestration is not moved into a CLI config but into a separate SDK - the Agent Development Kit - and that changes the very nature of the approach: here a system is less configured than assembled from building blocks, a coordinator and narrow sub-agents, the way an ordinary software project would be. Project memory is set by a GEMINI.md file, everyday settings by .gemini/settings.json, and agents talk to one another and to remote workers over the A2A protocol in the Antigravity environment. The model behind all of this is Gemini; Google, like the others, does not disclose its architecture.
What sets this environment apart is an explicit, deliberately built-in slider between two different modes of operation, rather than a one-time choice of either. At one end is a model-led mode, where the model itself decides which sub-agent to call and when, based on the current state of the task. At the other is a deterministic workflow, where the order of steps is fixed in advance and does not change depending on what the model "thought" at a given step. You can move between these extremes depending on how predictable a given process needs to be.
Technically, sub-agents in this environment run in one of three modes - chat, task or single-turn - and, unlike many other environments, can execute in parallel rather than strictly one after another. Antigravity adds another layer on top of this: remote subagents, which physically run not in the local process but on a remote side, and the Jules agent, connected over the same A2A protocol as ordinary sub-agents.
The official materials in this section form a sequence from the general to the specific: the ADK documentation explains how the coordinator and sub-agents are built as a programming model, the codelab on assembling a multi-agent system shows the same thing on a concrete working example, and the Gemini CLI and Antigravity docs cover what this looks like inside an actual shell and IDE.
The meta-prompt for Gemini sets GEMINI.md to the coordinator role: the agent reads the listed materials, writes in delegation rules for sub-agents, explicitly states the allowed chat, task and single-turn modes with the option of running in parallel, and fixes where model-led logic runs by default and where a strict deterministic workflow is required. The system is then actually assembled on ADK - as a coordinator plus a set of narrow sub-agents, not as a single markdown file of instructions.
As a separate step, the meta-prompt wires in remote subagents through Gemini CLI or Antigravity and the A2A link, if the task calls for it - and, as in the other environments, the whole plan is shown before changes are made, not after the fact.
Read:
- google.github.io/adk-docs
- github.com/google-gemini/gemini-cli
- antigravity.google/docs/agent
Set up THIS project for multi-agent work (ADK):
1. Create/update GEMINI.md: role - coordinator; delegation to
sub-agents; chat / task / single-turn modes (in parallel);
the model-led -> deterministic workflow slider.
2. Assemble the system on ADK: coordinator + narrow sub-agents.
3. Gemini CLI / Antigravity - remote subagents and A2A (Jules).
4. Balance quality vs token economy: narrow context,
summaries instead of dumps, no extra passes.
Show the plan BEFORE changes.