AgentRun
A workflow language for coding agents that mixes ordinary scripted steps with focused automated decisions and full agent calls, so a task is only handed to a model when it genuinely needs investigation.
AgentRun is a workflow domain-specific language for agents a developer already runs. A workflow defines repeatable steps, calls a lightweight decision-making component called Jev for a focused yes-or-no judgment, and calls a full agent only when a step actually needs investigation, while the surrounding application keeps its own tools, model access, permissions and budgets.
A bundled ticket-routing example demonstrates the pattern without needing an API key: for each incoming request, the workflow first tries a scripted tool search, asks Jev whether the resulting answer is good enough at a set confidence threshold, and if not, gives one attempt to a full agent before escalating the case for human review. The project reports how many agent calls and how many decision calls each case needed, and what the workflow decided.
It requires Node.js 22.19 or newer and is installed as an npm package, with an editor extension available for building workflows interactively.