← all projects

Qwen3 agent harness

Dimaag-AI · 2026

  • llm
  • agents
  • vlm

problem

an agent with many tools burns its context window on schemas and old turns before it does any work.

what I built

  • a Qwen3 32B agent harness with network-isolated code execution
  • LLM-based context compaction, and deduplicated tool schemas
  • document VLM and RAG ingest tools, and a benchmark of Qwen3-VL 32B and Qwen2.5-VL 72B

results

  • deduplicating tool schemas cut the fixed per-call cost 45% (9.2k to 5.0k tokens)
  • 85% of a 32k context window stays free

← all projects