Coding agent builder: the hard part is knowing when to act, not being smart — 82% on SWE-bench Verified
BarathKanna · reddit · 2026-10-24
A coding agent developer shares what took the most work: not making the agent smarter, but making it behave. The two annoying failure modes are the eager agent that refactors files you never asked about, and the timid one that asks "shall I proceed?" before every semicolon.
His fix is a context-aware split: act when the task is clear and contained; ask when there's real ambiguity or wide blast radius; never touch a defined set of things.
Supporting mechanisms:
- Self-verification: runs its work in a sandbox, finds and fixes bugs before reporting done
- Model routing: each prompt goes to the most efficient model automatically — painful to wire up but paid off in cost and performance
- Multimodal jobs: handles image generation too, not just text and code
It scores 82% (410/500) on SWE-bench Verified with the official harness, patches public. Trade-off: verification makes each task slower than pure generation, which he thinks is worth it. Built in India, free to use (disclosed as the author's own product).
More from coding & agent
- Microsoft open-sources Agensh, a decentralized multi-agent harness scaling to 1,024 agents — donglixp · 2026-10-24
- dumbtv strips Fire TV ads and tracking with one command, no root needed — Signalman23 · 2026-10-24
- CodeRabbit CLI v0.9.0 ships cr code for managing cloud agent tasks in terminal — leslysandra · 2026-10-24
- Solo dev turns a school timetable solver into an MCP server for Claude — Wilbek_DK · 2026-10-24
- ChatGPT Sites ships detailed site analytics: dev logs +216% weekly traffic — Dimillian · 2026-10-24
- Gemini Flash-Lite as a meeting copilot brain: real test results, what worked and what broke — DeskSignalAI · 2026-10-24