Developer open sources AgentRun DSL for converting agent work to workflows

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

On September 23, 2026, a developer at grep.ai open sourced AgentRun, a domain-specific language built to convert repeatable portions of AI agent work into structured, deterministic workflows. The tool draws on traces and retrospective notes that agents generate while completing jobs, using that record to identify which steps can be handled without invoking a full agent. Those steps are then expressed as combinations of tool calls, code, and what the developer describes as Jev-powered system-one decisions covering tasks like routing and evidence screening, with full agents reserved only for steps that genuinely require deeper investigation.

The practical motivation is cost. Running a full AI agent on every step of a job is expensive, and much of what agents do in practice turns out to be routine enough to be captured in a more rigid, cheaper form. AgentRun sits at a junction between fully autonomous agent pipelines and hand-written workflow automation, trying to let teams get the benefits of agent flexibility during initial exploration while progressively locking in the parts that have proven stable. The central question the tool raises is how reliably agent traces can be used to identify what is truly repeatable versus what only looked repeatable in the cases observed so far. That boundary between deterministic and non-deterministic work is genuinely contested in the broader field.

Because no reactions from the Pro-AI, Anti-AI, or Middle Ground camps have been published yet, their likely positions can only be anticipated. Pro-AI commentators would typically welcome a tool that makes agent deployments more efficient and cost-effective, framing it as evidence that the ecosystem is maturing. Anti-AI voices would likely raise concerns about the opacity of automatically generated workflows and the risk that errors baked into agent traces get institutionalized into production pipelines. The Middle Ground camp would probably focus on whether the retrospective note mechanism is robust enough to catch edge cases, and whether teams have meaningful visibility into the workflows AgentRun produces before those workflows run autonomously.

The key things to watch are community evaluations of AgentRun on real-world agent traces, particularly any reports on how often the DSL misclassifies a non-repeatable step as automatable, and whether grep.ai publishes further documentation on the Jev-powered decision layer that handles routing and screening.