SteP: Stacked LLM Policies for Web Actions

SteP is a web-agent framework that composes LLM policies through an explicit control stack rather than a single monolithic prompt. It evaluates on WebArena, MiniWoB++, and a CRM environment, and substantially improves WebArena performance over prior GPT-4-based baselines.

• ColorBrowserAgent: Complex Long-Horizon Browser Agent with Adaptive Knowledge Evolution
  January 12, 2026 · arXiv
• Inducing Programmatic Skills for Agentic Tasks
  April 9, 2025 · COLM 2025
• The Tool Illusion: Rethinking Tool Use in Web Agents
  April 3, 2026 · arXiv
• When Users Change Their Mind: Evaluating Interruptible Agents in Long-Horizon Web Navigation
  April 1, 2026 · arXiv
• WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at Scale
  March 2026 · Blog Post
• AI Planning Framework for LLM-Based Web Agents
  March 13, 2026 · arXiv