Loom – An event-driven OS for AI agents, built by a college junior in 10 days (github.com)

Loom is an event-driven runtime and “OS” for AI agents — an ambitious, production-minded framework (Apache 2.0) created in a short sprint by a college junior and released with docs, examples and a working voice demo. Instead of request/response plumbing, Loom centers on an EventBus: events in, stateful agents process and coordinate, and actions out. That architecture enables real‑time multi‑agent patterns (request/reply, fanout/fanin, contract‑net, barriers) with built‑in QoS tiers, backpressure strategies (sampling, dropping, aggregation), and a Router that applies policy-based Local/Cloud/Hybrid routing using privacy, latency, cost and quality budgets — useful for hybrid edge/cloud systems and safety-sensitive routing decisions. Technically, Loom’s core is Rust (loom-core) with vendored protobufs and a gRPC Bridge for polyglot agents; a Python SDK (loom-py) and JS SDK are supported/coming. Key primitives include a unified Envelope metadata model (thread_id, correlation_id, ttl, hop), capability registry (native Rust providers, MCP tools via JSON‑RPC stdio, WASM/out‑of‑process plugins), persistent agent state (RocksDB), optional VectorDB, OpenTelemetry/Prometheus telemetry, and audio stack (mic/VAD/STT/TTS) as an opt‑in module. The project ships collaboration primitives, timeout/error semantics, integration tests, and examples (voice_agent, trio). Loom’s combination of event semantics, capability federation (MCP), and policy routing makes it a compelling substrate for scalable, hybrid multi‑agent AI systems.