agisdk  by agi-inc

AI agent toolkit for real-world web environments

Created 1 year ago
419 stars

Top 69.6% on SourcePulse

GitHubView on GitHub
Project Summary

AGI SDK is a Python toolkit for building, evaluating, and benchmarking AI browser agents against real-world web applications. It targets AI researchers and developers seeking to test agent capabilities in realistic, complex environments, offering a standardized benchmark and a platform for comparing agent performance.

How It Works

The SDK provides high-fidelity, deterministic web application clones (e.g., Amazon, DoorDash) built with modern web stacks. It uses a harness that orchestrates agent interactions with these environments, providing structured observations (DOM, accessibility tree, screenshots) and accepting actions as function calls. This approach allows for reproducible evaluation and direct comparison of agents on standardized tasks.

Quick Start & Requirements

Highlighted Details

  • Supports OpenAI, Anthropic, and OpenRouter models.
  • Includes full-stack web replicas of popular applications like Amazon, DoorDash, Airbnb, Gmail, and LinkedIn.
  • Features a leaderboard (REAL Bench) for agent performance comparison.
  • Offers a robust agent API with observations, actions, memory, and error handling.
  • Allows customization of the harness and integration of custom agents.

Maintenance & Community

  • Community channels are listed as "coming soon" on Discord.
  • AGI Inc. can be followed on LinkedIn.

Licensing & Compatibility

  • The repository does not explicitly state a license in the provided README.

Limitations & Caveats

  • Discord community channels are not yet available.
  • The README does not specify licensing details, which may impact commercial use or closed-source integration.
Health Check
Last Commit

7 months ago

Responsiveness

Inactive

Pull Requests (30d)
0
Issues (30d)
1
Star History
8 stars in the last 30 days

Explore Similar Projects

Starred by Shizhe Diao Shizhe Diao(Author of LMFlow; Research Scientist at NVIDIA), Michael Chiang Michael Chiang(Cofounder of Ollama), and
7 more.

openbench by groq

0%
784
Provider-agnostic LLM evaluation infrastructure
Created 11 months ago
Updated 2 weeks ago
Feedback? Help us improve.