Kon pairs Qwen with tiny harness

This is less a benchmark than a useful signal: local coding agents are starting to compete on harness overhead, not just model quality.

• –Kon’s small prompt and six-tool core make it attractive for local models where every fixed token cuts into repo context.
• –The Qwen3.6-27B pairing matters because dense mid-size models are becoming plausible workers for codebase search, triage, and subagent loops.
• –Recent community additions like LaTeX, permissions, web search, and fetch tools suggest Kon is maturing without abandoning its “small core” premise.
• –The obvious caveat: this is a one-user field report, not a repeatable coding-agent benchmark against Claude Code, OpenCode, or mini-swe-agent.