Dev's custom harness lets Qwen 27B autonomously build a C99 compiler for x64 ELF

Naiw80 · reddit · 2026-08-25

A developer shares a 6-month project: building a custom agent harness around Qwen 3.6 (later 3.8) 27B, first fixing small-model failure modes like repetition and empty replies, then adding a planner, coder, debugger, researcher and validators as dedicated subagents, forcing a research-then-execute pattern instead of relying on memorized knowledge.

The prompt: "build a C99 compiler producing working x64 ELFs." Hardware: Tesla P100 + RTX 4070 (13-14 tok/s), with Gemma4 12B on an Arc B580 acting as validator.

Key lessons:

Context management was the hardest part: with llama.cpp lacking context-shift, bad compaction/pruning triggered repeated 6-7 min prefills; solving this took 3 weeks and is the main reason the project took 6 weeks.
x86 codegen: the model hallucinated opcodes instead of looking them up; rigid system prompts and mandating libcapstone largely fixed it.

Longest uninterrupted run: 1 week. Conclusion: Qwen 27B is extremely capable when steered properly. Output is open-sourced (github.com/Na1w/tc); the orchestrator is not yet released.

Original post →

Dev's custom harness lets Qwen 27B autonomously build a C99 compiler for x64 ELF

More from coding & agent