Aulay is an agent platform that mixes hosted cloud agents, local CLI tools, and deterministic workflows in a single session — with resumable state, audit trails, and a pluggable runtime contract.
The thesis
“The future of AI agents is deterministic frames between non-deterministic loads.”
LLM calls, tool runs, web fetches, code execution — the opaque, stochastic, hard-to-debug parts of an agent.
Graphs, state machines, sandboxes, queues — the deterministic structure around the loads that makes an agent’s behaviour reproducible and auditable.
Most agent frameworks today are all load — one giant LLM loop with no frame around it. Aulay inverts this: a hub where multiple runtimes live side-by-side. Some contribute frames. Some contribute loads. The frame runtimes call into the load runtimes at well-defined nodes.
Architecture
Strict dependency rule: Aulay OS depends on Fabric; Fabric never depends on Aulay OS. Every seam is a pluggable interface, so you can swap storage, runtimes, or the host application without touching the core.
Layer 1
The application platform. HTTP API, web app runtime, event bus, automation engine, resource runtime, and the IDE / Chat / Admin / Files frontends.
Layer 2
The agent hub. Owns runtime control plane, sandbox service, agent package catalog, and domain package loader. Runs standalone; hosts agents from any source.
Layer 3
Hosts the actual agent processes. Runs locally, in Docker, in Kubernetes, or on a remote machine. Stateless from its own point of view — all persistent state lives in Fabric.
Runtimes
A Fabric runtime plug-in is a separate process that speaks a small WebSocket protocol. No shared library, no FFI, no dylibs. Any language works. Three canonical plug-ins ship in-tree and exercise the contract end-to-end.
aulay.runtimeMany small loads
Local CLI tool harness. Hosts claude-code, cursor-agent, codex, opencode, pi, terminal, and git as first-class agent backends.
cursor.cloudRemote big load
Hosted background agents via Cursor's cloud API. Pure WebSocket client of fabric-daemon; no ports, no host-side binaries.
aulay.flowThe frame provider
Aulay's deterministic graph-flow runtime. Graphs of transforms, branches, and sub-agent invocations that call into load plug-ins at well-defined nodes.
Automation
AutomationFlow wraps your agents in graph-shaped workflows with durable state, approval gates, and child-agent orchestration. Long jobs survive restarts. Humans can step in without breaking the trace.
Every workflow run is persisted. Nothing is lost to a crash or a restart — pick up from the last checkpoint.
Pause a run mid-flight to collect human input. Resume exactly where it stopped, with full context intact.
A parent agent can spawn and supervise children across runtimes, collect their results, and branch on them.
Every load, every frame transition, every decision is a record. Re-play the whole run, or diff two runs side-by-side.
Apps
Every surface in Aulay is a microapp — registered by manifest, served by the app runtime, sharing the same auth and session model. You can swap any of them out, or ship your own alongside them.
Code editor with agents, PTYs, and inline markdown streams.
Conversational interface that hosts agents from any runtime.
Platform administration, graph visualisation, YAML editing.
PTY-backed terminal, wired to sessions and agents.
File browser + editor with markdown preview.
Discover agent types and inspect their graphs.
Author flows, trigger runs, follow live execution.
Cron-style scheduling for workflows and agents.
Messages and notifications across sessions.
Connect Aulay sessions into n8n automations.