Harness engineering masks the real issue: agents lack compositionality
qi2peng2 · x · 2026-10-13
Researcher qi2peng2 argues that the harness engineering boom is masking a deeper problem with agents: neural models are at best simulators of symbolic processes, sensitive to variable names, inputs, and convoluted control flows. Placing them in the outermost loop of agents yields a mountain of text and code with no behavioral guarantees and poor editability — a lack of the compositionality that once made NLP elegant.
More from AGI Musings
- David Patterson: all cars will be autonomous by 2030, saving 40,000 American lives a year — davidpattersonx · 2026-10-13
- A more competent alt-rat leader could reshape the rationalist community's dynamics — jessi_cata · 2026-10-13
- Dan Shipper: cracks are appearing in the frontier model duopoly — danshipper · 2026-10-13
- Stop chasing AI frameworks: dev says spend 6 months on the math instead — TivadarDanka · 2026-10-13
- Anthropic researcher recounts conversion: frontier insiders converge on same AGI conclusions — abhiadesai · 2026-10-13
- Esther Dyson's new book 'In Defense of Limits' asks how humans should govern AI — Grady_Booch · 2026-10-13