Lilian Weng highlights harness engineering as a path to recursive self-improvement in AI systems

Lilian Weng, cofounder at Thinky, published a synthesis of 35 papers on harness engineering for recursive self-improvement (RSI), arguing that harnesses—not direct weight updates—are likely to be the dominant mechanism for future AI self-improvement. In her post, she frames RSI as a feedback loop where models improve the training pipeline and deployment systems that, in turn, enable better successor models. She also emphasizes that even as improvements are internalized into core models, the need to specify goals and context will persist.

The synthesis recaps proven design trends in harnesses and surveys recent literature, highlighting the ACE paper and newer work such as Meta-Harnesses. Weng suggests harness engineering will evolve toward self-improvement and enable auto-research, positioning harnesses as a foundational layer for smarter systems.

Industry product announcements this week reflect the same shift. Anthropic expanded its "background agent" user experience across mobile and web with Claude Cowork, positioning Claude as a task-running background teammate rather than a foreground chat interface. Google added background execution, remote MCP servers, custom function calling, and credential refresh to the Gemini API Managed Agents. LangChain launched a Deep Agents course and an open-source harness project, while Hermes Agent added pluggable secrets managers and native 1Password integration.

Meta previewed Muse Video alongside the earlier Muse Image release, emphasizing an explicitly agentic generation loop: planning, web search, tool use, code execution, and self-refinement before rendering. Meta reports that performance improves with scaled test-time compute and that self-refinement behavior emerged during reinforcement learning rather than being hand-scripted.

NVIDIA and Cohere also shipped audio releases. NVIDIA released Audex, a 30B parameter/3B active MoE with 1M context for unified text+audio work, claiming preservation of text intelligence alongside broad audio generation and understanding via a single MoE backbone. Cohere launched Cohere Transcribe Arabic as an Apache 2.0–licensed ASR model focused on dialects, code-switching, and Arabic-accented English.