LoopEngineBETA

Docs

Documentation

How LoopEngine works, from scaffold to durable approvals — agents, adapters, permission rules, and abilities.

Discovery & Registry

core/discover-agents.ts and core/agent-registry.ts are the convention-over-configuration layer: an agents/<name>/ folder with optional tools/, skills/, and actauth.yml is all a new agent needs (see Directory Structure and Add an Agent) — no registration step to remember. Every adapter resolves "which agent" through the registry; it scans agents/ and keys each entry by AgentConfig.name, not the folder name, at import time.

agents/<name>/subagents/* is a separate folder convention on top of this: drop another agent's folder there and it becomes one of this agent's tools automatically (see Tools's agentAsTool), merged into tools on top of whatever the agent's own config resolves to.

Worth being explicit about the boundary this points at: agents/ in the loopengine repo itself — customer-service, file-agent, rag-agent, support-orchestrator, web-agent — is that repo's own example app, not part of the published package surface. The actual public API is index.ts: runAgent, AgentConfig types, the full LoopEvent union, client.ts's functions, SessionStore/SessionKnit, SkillGarden — everything a consumer building on loopengine is meant to import, re-exported through one entry point since the package has no exports map.