Microsoft's BoxBox lets agents hot-reload their own harness mid-run, hitting 1.70x speedup
dair_ai · x · 2026-11-10
dair-ai highlights BoxBox, a Microsoft-led paper (with Nanjing/Zhejiang/Birkbeck) on harness self-improvement for long-running agents. Offline harness tuning needs reference answers; open-ended multi-hour tasks don't have them. BoxBox moves optimization into the task itself: the agent decides when to revise its harness, and a kernel hot-reloads it while preserving execution context — no restart. On PerfRace, a new benchmark where agents spend hours speeding up real software projects, BoxBox reaches Speedup@4h of 1.70x vs 1.54x for the strongest baseline, at lower cost ($16.41 vs $23+).
More from coding & agent
- Claude Fable 5.2 coordinates 1,000+ subagents for 8 days to boot a working OS prototype — kimmonismus · 2026-11-10
- NODES 2026: 100+ sessions, but devs should filter for Text2Cypher, entity resolution and quantization — JeremyCMorgan · 2026-11-10
- Astral open-sources serc under new astral-experiments org for agent-authored projects — charliermarsh · 2026-11-10
- Astral's serc: Codex-built Rust CPython bytecode compiler, byte-identical and 2-4x faster — charliermarsh · 2026-11-10
- Sentry CEO shares hands-free AI game dev progress: auto-added wildlife shaders at 82fps — zeeg · 2026-11-10
- Anthropic's official guide: when to use Fable 5.2, Opus 5.5, Sonnet or Haiku — ClaudeDevs · 2026-11-10