Practitioner: novelty search with agents is still extremely hard, but reveals a big capability overhang
generativist · x · 2026-11-20
An AI practitioner shares first-hand observations on running novelty search with agents:
- It remains extraordinarily hard — "those basins really really bind": the search gets trapped in local attractor basins and struggles to escape existing patterns
- Yet even so, agents keep surfacing things novel to him, in domains he knows well, without any literature search — evidence of a significant capability overhang: models already possess far more capacity than current applications exploit
A rare first-hand verdict on exploratory agent workflows, highlighting the gap between method bottlenecks and latent model capability.
More from coding & agent
- 200 lines of Python + Gemini generates a synced-caption, voice-cloned walkthrough video — DynamicWebPaige · 2026-11-20
- Strategist evolves reasoning strategies as code, wins 24/30 benchmarks at 39% lower cost — liuzhuang1234 · 2026-11-20
- RestTemplate is deprecated in Spring Boot 4.2 — here's the direct swap to RestClient — therealdanvega · 2026-11-20
- How a 50-customer B2B app automated QA with agents simulated from production MLflow traces — Lopsided_Feedback_81 · 2026-11-20
- Developer gives Claude Code its own computer to drive browsers, spreadsheets and desktop apps — jradoff · 2026-11-20
- Garry Tan shares GBrain cases: AI working overnight, 850 pages in 3 weeks — garrytan · 2026-11-20