Most agent frameworks let a model call whatever it likes and tell you the cost afterwards. Loom flips that. Work is declared as a dataflow of stages (ordinary transforms plus AI steps such as infer and summarize) and compiled into a plan, and the plan can say what it will cost before a single model call is made. When it runs, every stage is admitted against rate limits and a run-level dollar budget, failures are classified and recovered by kind, and every task carries only the credentials it was granted.
Four layers. Authoring: a Go API and a visual studio that exports to Go. Compile: validate the graph, fuse what can be fused, fingerprint every step, and emit minimal task envelopes. Run: a scheduler that governs budgets and retries and executes either in-process or through durable workers on a leased queue. Foundations shared by both: content-addressed storage that doubles as checkpoints and lineage, a security broker for credentials and egress with an audit log, a model registry across providers, an MCP client for tools, and an event stream that a live visualizer draws as a constellation.