Agent Composition
Xema has one Agent primitive. A leaf specialist is an Agent with no subagents; a multi-agent composition is the same Agent shape with a recursive subagent tree. The same published Agent powers Interactive Sessions and Workflow steps — there is no separate session-Agent, workflow-Agent, kernel-Agent, or composition storage model.
Agent Composition replaces the old single-agent-per-phase model. The Model Resolution Matrix picks the right model for each node at invocation time based on declared dimensions — no hardcoded phase-to-model mapping.
Quick Links
| Page | What it covers |
|---|---|
| Concepts | The unified Agent, inheritance, recursion, revisions, and execution experiences |
| Model Resolution | ModelResolutionRule, dimensions, priority, debugging |
Getting Started
- Concepts — understand the unified Agent, recursive nodes, and revision lifecycle.
- Model Resolution — configure which model each node uses.
- Build and publish Agents in Agent Studio or through the Agent authoring API.
FAQ
Q: Can I use Agent Composition for both interactive sessions and automated workflows? A: Yes. The same published Agent can be launched as an interactive session or as a workflow step. The launch surface changes the interaction and lifecycle behavior, not the Agent definition.
Q: What happens if an Agent references an unpublished subagent? A: Published resolution fails fast rather than silently selecting a draft. Draft resolution is limited to the explicit authoring sandbox.
Q: Can the model change mid-turn?
A: No. The Model Resolution Matrix resolves the model at invocation boundaries only: agent start, sub-agent spawn, and /skill launch. The model is fixed for the duration of a turn.