An open standard for autonomous agents. Ingest signals, score decisions, execute actions. One primitive. One audit trail. One conformance test.
STORM is an open, neutral standard for building autonomous agents. It collapses the gap between signal and action by exposing a single primitive — one that ingests events, scores them against a policy, and executes the resulting action with adaptive sizing and full replayability.
This document specifies the protocol, the runtime contract, and the conformance rules for any STORM-compatible agent.
Agentic software today is fragmented across orchestrators, frameworks, and bespoke glue. The result: opaque cycles, untraceable decisions, and lock-in to a particular vendor's stack.
STORM exists to make agentic software boring. A single primitive. A single audit trail. A single conformance test. No orchestrator required.
A STORM cycle is a tuple (signal, score, action, log). Signals are typed events from any source — HTTP, webhooks, streams, model outputs. Scores are deterministic functions over signals. Actions are side-effects gated by score thresholds. Logs are immutable records of every cycle, replayable bit-for-bit.
Conformance requires that any two implementations, given the same signals and policy, produce the same scores and the same action queue.
The runtime exposes four hooks: ingest, score, execute, log. Implementations may add memory, sizing, and observability layers, but must preserve the cycle invariant: every executed action must be reproducible from its log.
Sizing is adaptive but bounded. Memory is persistent by default. Observability is structured and machine-readable.
STORM is MIT licensed and stewarded by an open working group. The reference implementation lives on GitHub. Spec changes follow a public RFC process.
No single vendor controls the standard. Conformance tests are open and free to run.