The framework-agnostic agent layer
@robbiesrobotics/alice-agents defines what an agent is. Workspace-file convention with SOUL.md, AGENTS.md, MEMORY.md, and a 31-persona registry. Runtimes read these files.
What makes an agent
A.L.I.C.E. agents aren't just prompts — they're persistent, structured entities with defined roles, memory, learnings, and feedback loops. The agent layer defines a filesystem convention that any runtime can read and execute against.
This is the I8 boundary in Hub's packages: the agent definitions are runtime-agnostic, so alice-runtime and OpenClaw can both execute the same agent workspaces. Swap the runtime without rewriting your agents.
The 31-persona registry covers the full spectrum of business and technical domains — from frontend and backend engineering to HR, legal, finance, and autonomous research. agents-starter.json and agents-pro.json provide different scale configurations for different team sizes.
The workspace file system
SOUL.md
Agent persona, values, and behavioral boundaries
AGENTS.md
Team structure, roles, and delegation rules
MEMORY.md
Semantic memory, facts, and learned knowledge
LEARNINGS.md
Session lessons, patterns, and improvement notes
PLAYBOOK.md
Domain expertise, proven patterns, and playbooks
FEEDBACK.md
User and peer feedback for iteration
31-persona registry
agents-starter.json and agents-pro.json provide two registry sizes. The full 31-persona registry covers:
The I8 boundary
Stack flow: Hub routes a request → picks a framework → calls Runtime over OpenAI-compat HTTP → Runtime loads the Agents workspace + persona → tool-calls Memory via FFI.
Define your agents once.
Run them anywhere — OpenClaw, alice-runtime, or any OpenAI-compatible runtime. The agent layer is portable by design.