Harness routing — auto-pick cheap vs frontier models per task

A router that sends easy coding-agent tasks to cheap models and hard ones to frontier models, with a semantic layer that keeps context correct and costs down.

Problem

Heavy coding-agent users notice they send every request to the same model tier — wasteful and unnecessary. Some already build their own four-tier routing (easy to cheap models, hard to frontier) and want more savings layers, but these harnesses are one-off and lack a semantic layer to keep queries and code context correct.

Solution

A CLI/local daemon that classifies task difficulty and routes across easy/medium/hard/frontier tiers, ships a token-efficient semantic context layer to keep code and queries grounded, and logs per-tier cost and success so users can replay, tune, and export a shareable harness config.

Users

Developers and small teams who heavily use coding agents (Claude Code, Opencode, Cursor, etc.).

MVP

Rule-based difficulty classifier v1; support for 2-4 model providers; per-tier cost and latency dashboard; save/load harness configs; automatic fallback to a higher tier on failure

Revenue

Open-core with a hosted dashboard subscription; usage-based proxy billing; or a team plan for shared cost policies.

Why now

Multiple Show HN harness projects signal real demand and early adopters; widening price gaps between models make routing immediately and measurably money-saving.