Under matched token budgets, web agent memory and skill augmentations lose much of their edge
MMedAgent · hf · 2026-10-23
New research (arXiv:2606.15017) by gspandana et al. asks whether test-time online augmentations — memory, skills, workflows — still help web agents when total inference tokens are controlled.
- Methods like AWM, ASI and ReasoningBank do help, but burn many more inference tokens.
- The authors re-evaluated all three on WebArena against a vanilla agent given the same total budget, free to spend it on more interaction steps.
- Under matched budgets, much of the reported advantage of all three methods disappears.
- Takeaways: compare agents at equal token count, not equal step count; scaffolding for its own sake yields poor performance per token — a result practitioners say they also see in production.
Related event: Web Agent Online Enhancements Lose Most Gains Under Equal Token Budgets(2 posts)→
More from coding & agent
- Richard Socher's Recursive launches Flint, a fast local coding model built in under 4 months — RichardSocher · 2026-10-23
- Peking University and Alibaba unveil OmniDR-Bench for video-centered deep research agents — Qwen · 2026-10-23
- $100K Open Kernel Challenge accused of copying GPU MODE OSS code via AI agent rewrite — josh_wills · 2026-10-23
- GBrain Adds Multi-Account Switching So Your AI Memory Follows Each Identity — garrytan · 2026-10-23
- Open textbook Kubernetes System Design in Small Easy Steps lands on ChapterPal: 78 pages, 19 exercises — burkov · 2026-10-23
- WooCommerce MCP server lets Claude query store orders, stock and sales (read-only by default) — thejester1324 · 2026-10-23