HERA: co-evolving agent harnesses and environments for reliable abstention
UW · hf · 2026-10-20
HERA (Harness-Environment Co-Evolution) co-evolves agent harnesses with task environments to improve reliable abstention—knowing when to decline rather than produce wrong answers. The method works across diverse LLMs without model-specific tuning, improving both abstention quality and task completion rates.
More from coding & agent
- Factory adds Linear delegations, letting Droid agents run issues automatically — matanSF · 2026-10-21
- Agent memory, goals and emotion should be designed as one system, argues dev — mimi10v3 · 2026-10-21
- Meta launches Muse Wonder coding models with 1M context, $4/$20 per million tokens — AIatMeta · 2026-10-21
- CastWorkshop: open-source studio ties Ollama and ComfyUI into persistent AI characters — CastWorkshop · 2026-10-21
- Agents break in production for boring backend reasons, not model reasons — DayKnown1050 · 2026-10-21
- REA turns any site, app, or game into an RL environment via open-source AgentEnv — henloitsjoyce · 2026-10-21