Homepage | Github Repo

Remoraflow is a JSON-based DSL where AI agents define, compile, and execute reliable and consistent workflows.

AI workflows have historically been defined by the language LLMs know best: prompts. However, prompts aren’t functions; they can’t guarantee a particular input or output space, can’t justify their control flow, and can’t provide the privacy, reliability, or auditability demanded by any serious business. Remoraflow solves this by giving agents a standard language for defining, validating, and executing semi-deterministic AI workflows.

The language is purpose-built for agents to author, not just execute. Workflows are defined as a graph of typed steps that pass data through a shared scope via JMESPath expressions. Control flow is fixed, steps are type-safe, and an LLM is invoked only at the steps where one is explicitly placed, so genuine intelligence is applied where it’s needed and everything else stays consistent. Agents can generate these workflows through an ordinary tool call, and a bundled viewer lets humans inspect and edit the resulting graph or audit a previous execution.

Before a workflow ever runs, a compiler statically analyzes it, catching broken references, type mismatches, and unreachable steps to cut down on runtime failures and wasted LLM API calls. The compiler also distinguishes static tool parameters from dynamic ones, letting supervisors review and approve a workflow’s behavioral envelope ahead of time.

For teams, RemoraFlow adds the controls organizations need to trust agentic systems: approval routing, timeouts, and a full audit trail, with tiered policies for different classes of action. Compiled workflows run anywhere on a durable execution engine—surviving restarts and resuming cleanly from the point of interruption.