Durable agents as infrastructure.
In Norse mythology, the Norns weave fate — shaping what happens next based on what has come before. A Norn agent plans, executes, adapts, and survives. It is not a tool you prompt. It is infrastructure you deploy.
What is Norn
The AI agent engine. It gives agents the ability to read and write files, run commands, search code, and coordinate with other agents — designed to run on the full Ablative stack.
The problem with AI agents today
Most agent frameworks treat agents as scripts that call an LLM in a loop. They run in a single process, lose state on crash, and have no meaningful supervision. If the process dies, the agent dies. If the agent hallucinates, nobody catches it. If it needs to wait three days for a human approval, it either polls or gives up.
Norn is designed to treat agents as infrastructure — a durable workflow on Aion, a communicating process on beamr, with state in haematite and coordination through liminal. Kill the server — the agent resumes. Deploy a new version — running agents keep their version. An agent that sleeps for a week costs nothing until it wakes up.
What makes a Norn agent different
Durable by design
Agent steps are designed to be event-sourced through Aion. Crash, restart, replay. The agent picks up exactly where it left off — no checkpointing, no manual save/restore. Durability is not a feature. It is the target execution model.
Supervised processes
Agents are designed to run as BEAM processes on beamr with supervision trees. Sub-tasks crash-isolate. A failing tool call doesn't take down the agent — the supervisor restarts the failed child and the agent continues.
Multi-agent coordination
Agents are designed to communicate through Liminal channels and coordinate through Aion workflows. A team of agents working on a shared problem is not a hack — it is a first-class orchestration pattern with state management and crash isolation.
Observable by design
Every decision, tool call, and state transition is recorded. Audit an agent's reasoning after the fact. Replay a decision point with different context. The agent shows its work because the architecture requires it.
The north star
Deploy an agent the way you deploy a service. Give it a name, a purpose, and the tools it needs. It runs until the work is done — whether that takes five seconds or five months. It survives infrastructure failures, shows its reasoning, and coordinates with other agents through typed channels. That is the promise of durable agents as infrastructure.
AGPL-3.0 licensed. View on GitHub →