Rebuilding agent skill retrieval and model routing with Jev
After testing 145 custom skills, the author found Opus 5 took 30 seconds to find the right tool while Jev hit it in 5 seconds and cut wrong picks by more than half, so they switched their routing backbone to Jev, bringing initialization to sub-second and cutting API costs by 78%.
If your coding agent takes 30 seconds just to search its own skill library and pick an LLM, your agent harness is already hopelessly broken Jay broke down why developers with dozens of skills are hitting a wall in Claude Code and Codex In a test across 145 custom skills, Claude Opus 5 burned 30 seconds just locating the right tool, while Jev retrieved the exact skill in 5 seconds flat while cutting wrong picks by more than half Even OpenAI just rushed out their Decisions API because they know decision models are replacing prompt-based tool calling I tore down our slow agent orchestrator and rebuilt our routing backbone around Jev Instead of paying frontier models to deliberate over file paths and skill lookups, Jev reads incoming user prompts in a third of a second and handles model routing instantly Here is how I wired the Jev routing engine into our developer workflows: step 1 - turn your skill registry into a structured menu of choices instead of stuffing full markdown files into your prompt step 2 - run a fast Jev Choice check that matches user intent to the exact skill in milliseconds step 3 - configure an automated model router that assigns trivial lookups to Haiku and reserves Opus exclusively for deep architectural refactoring step 4 - calibrate confidence thresholds so ambiguous requests trigger human clarification before running destructive terminal commands step 5 - build a validation set of 100 historical tasks to benchmark Jev skill selections and fine-tune criteria The shift radically supercharged our software engineering velocity: agent task initialization dropped from 30 seconds to sub-second speed, daily API costs plummeted by 78%, and our developers execute complex multi-skill workflows without hitting execution timeouts We turned a sluggish agent loop into an instant, high-precision development environment Stop wasting frontier reasoning models on tool selection and build a dedicated Jev decision backbone before the rest of the industry leaves you behind
This post quotes another post: x.com/mn0g0znalll/status/2105292921591226722