Skip to content

Research

lewistham9x edited this page Mar 7, 2026 · 1 revision

Research

Academic papers that validate unbrowse's core approach: AI agents perform better with API access than browser automation.

API-Based Agents Outperform Browser Agents

"Beyond Browsing: API-Based Web Agents"

  • arXiv: 2410.16464
  • Authors: Yueqi Song et al. (Oct 2024, updated Jan 2025)
  • Key Finding: API-based agents outperform browser-based agents. Hybrid agents (API + browser) achieve 38.9% success rate on WebArena — a 24% absolute improvement over browsing alone (14.9%).

"AutoScraper: A Progressive Understanding Web Agent for Web Scraper Generation"

  • arXiv: 2404.12753
  • Venue: EMNLP 2024
  • Key Finding: LLMs can generate web scrapers automatically, but approaches suffer from limited adaptability. Two-stage framework proposed for diverse web environments. Unbrowse goes further — capturing actual internal API calls rather than generating HTML parsers.

"A Survey of WebAgents: Towards Next-Generation AI Agents"

  • arXiv: 2503.23350
  • Venue: KDD 2025
  • Key Finding: Comprehensive survey covering AutoGPT, BrowserAgent, and others. Identifies trust, safety, privacy, and generalizability as key challenges. API-based interaction is identified as a promising research direction.

Browser Agent Security Risks

"The Hidden Dangers of Browsing AI Agents"

  • arXiv: 2505.13076
  • Date: May 2025
  • Key Finding: Browsing agents have significant security vulnerabilities — prompt injection, credential exfiltration, domain bypass. CVE disclosed. API-based interaction avoids many of these attack surfaces.

The Agent Economy

"The Agent Economy: A Blockchain-Based Foundation for Autonomous AI Agents"

  • arXiv: 2602.14219
  • Date: February 2026
  • Key Finding: Proposes "Internet of Agents" with blockchain for permissionless participation, trustless settlement, and machine-to-machine micropayments. 5-layer architecture including MCP for tooling.

"Towards Multi-Agent Economies: A2A + x402 Micropayments"

  • arXiv: 2507.19550
  • Date: July 2025
  • Key Finding: Architecture combining A2A protocol with DLT-based agent discovery and x402 HTTP 402 micropayments. Enables agents to discover, authenticate, and compensate each other.

"Secure Autonomous Agent Payments"

  • arXiv: 2511.15712
  • Date: November 2025
  • Key Finding: Blockchain framework for verifying AI agent transaction authenticity and intent using DIDs, ZKPs, and TEE attestations.

"Can We Govern the Agent-to-Agent Economy?"

  • arXiv: 2501.16606
  • Date: January 2025
  • Key Finding: Explores governance of AI agent value exchange. Examines smart contracts and crypto as transactional substrate.

The Agentic Web

"Agentic Web: Weaving the Next Web with AI Agents"

  • arXiv: 2507.21206
  • Authors: Include Pieter Abbeel and Dawn Song (UC Berkeley)
  • Date: July 2025
  • Key Finding: Defines the "Agentic Web" — the next phase of the internet with autonomous, goal-driven agent interactions. Introduces the concept of an Agent Attention Economy.

"What Is Your AI Agent Buying?"

  • arXiv: 2508.02630
  • Date: August 2025
  • Key Finding: AI agents in e-commerce show choice homogeneity, position biases, and instability across model updates. Agent-mediated commerce behaves fundamentally differently from human markets.

Key Takeaways

  1. API > Browser is empirically proven — 24% improvement in success rates
  2. Browser agents are security risks — prompt injection and credential exfiltration vulnerabilities
  3. Agent economies are being formalized in academia with micropayment architectures
  4. The Agentic Web is recognized by top researchers as the next phase of the internet

The Gap Unbrowse Fills

Existing research assumes APIs are pre-existing. Unbrowse creates API access where none is formally provided — by capturing and packaging the internal APIs that power every website's frontend.