How this coding agent builder works daily: auto model routing, self-checking, 82% on SWE-bench

BarathKanna · reddit · 2026-10-24

A coding agent builder (Oxcode) shares his practical daily workflow:

No manual model picking: the agent routes each prompt to the most efficient model, beating a single fixed model on both cost and performance.
Delegate whole tasks, not next lines: e.g. "add pagination and update the tests" — the agent decides whether to proceed, ask questions, or leave files alone.
Self-checking before review: it runs changes in a sandbox, finds and fixes its own bugs before handing over the diff — slower per task but less debugging later.
Multimodal jobs, same flow: image generation goes through the same pipeline as code.
Verifiable results: 82% (410/500) on SWE-bench Verified with the official harness; all patches are public at github.com/Oxcode-ai/oxcode-evals.

He also asks whether people still hand-pick models per task.

More from coding & agent