Grep.ai AgentRun — enterprise harness with typed Jev nodes
On 18 September 2026, Grep.ai CTO Miguel Ríos published the X article AgentRun: a harness for repetitive knowledge work (product surface: grep.ai).
AgentRun is framed as a workflow harness, not a new decision model. A frontier agent does an expensive run once, leaves notes, then authors a program in the AgentRun DSL. Ordinary cases then run as inspectable steps — research / judge / report / operator agents, plain code, and typed Jev nodes (Route, Classify, Sift, Pick, plus verify clauses and confidence gates). The article positions AgentRun as “the first harness built around Jev” — treat that ranking claim as UnverifiedClaim.
The long AML alert walkthrough quotes large cost and latency wins versus a full frontier agent (e.g. ~$2.89 → ~$0.25 per alert on a 100-alert arm; broader 100k projections). Those figures are vendor-reported UnverifiedClaim, not ModelSystem.One measurements. Governance angle in the piece: each Jev answer keeps probabilities so judgment can be replayed without re-running research.
Primary sources: X post + article · grep.ai.