Movez Publishes a 10-Step Blueprint for Building a Jev Harness for Coding Agents

Summary

Movez founder Diogo Amogo has released a 12-page PDF describing a 10-step blueprint for building a Jev Harness around coding agents. The post claims the approach could make agents 200 times faster and 400 times cheaper, but those figures are presented as the blueprint’s claim rather than as independently reported measurements. In the proposed design, the LLM writes while the harness executes actions and Jev decides what each turn sees, where work is routed, and whether it runs. The blueprint warns that switching between models can be more expensive than using one model when the full context must be reprocessed, citing a comparison of $6.19 for an Opus-Sonnet-Opus route versus $4.15 for pure Opus. It says reading and searching account for 56.2% of tool turns and 46.5% of tokens, while code writing uses less than 10%, and recommends scoring retrieved chunks per query and compressing them after the question is known. Other recommendations include tiered tool disclosure, conditional loading of project instructions, routing sensitive work to first-party frontier models, sharing one retrieval pass across background tasks, and enforcing programmable allow, ask, or deny policies that inspect scripts before execution.