Agent Slots

An Agent Slot is an evolvable interaction boundary in an Agent Core: it lets Core-defined behavior enter the agent loop at governed points and compose tools, skills, MCP, state, or other agents without changing the host harness.

The defining property of a slot is not the capability it provides. It is the place where a capability can affect the agent loop under host governance. The host still owns scheduling, provider calls, tool dispatch, approvals, delivery, state enforcement, promotion, and rollback.

Slots are separate from the things they compose:

• A tool is a model-callable action.
• A skill is reusable knowledge, workflow, or policy.
• MCP is a protocol for external tools and context.
• An agent is another runnable loop.
• A package is a distribution unit that can install slots, tools, skills, libraries, and child cores together.

Current Demiurge slots are bootstrap, input, and output slots. They let an Agent Core add session context, shape current-turn input, handle final output, and compose tools, skills, MCP, state, or other agents through governed interfaces. Future slot kinds should represent new governed interaction boundaries, not ordinary feature categories.

This makes slots a natural evolution surface. A proposal worktree can replace, reorder, or compose slot behavior as files, while the host keeps risky effects behind stable contracts.