OpenSage: Next-generation Agentic AI

A self-contained ecosystem for constructing next-generation AI systems that enable AI to create and evolve agents, with an agentic-native training and inference framework.

Overview

OpenSage is a self-contained ecosystem for building next-generation AI systems that can self-evolve. It consists of four core components that together enable an end-to-end pipeline for agent construction, deployment, training, and inference.

  • OpenSage-ADK: An AI-centric Agent Development Kit that enables AI to self-create agent topologies and synthesize toolsets, rather than relying on hand-crafted agent designs.
  • OpenSage-Inference: AGInfer, an LLM inference engine optimized for agentic tasks, featuring long-context and tool-calling support with a modular design for high extensibility across new model architectures.
  • OpenSage-Training: A framework for large-scale agentic RL training with seamless integration into popular training frameworks such as AReaL and SLIME, built on the OpenSage agent scaffold and environment infrastructure.
  • OpenSage-Env: Integrated with popular environment frameworks (e.g., OpenEnv), with all agent execution running on NitroBox — an OpenSage-developed sandbox environment designed for efficient agentic RL training.

Highlights

OpenSage-ADK

SageAgent, built by AI itself using OpenSage-ADK, significantly outperforms other open-source agents and achieves state-of-the-art results across software engineering and cybersecurity benchmarks.

CyberGym
60.2%
vs. 39.4% OpenHands
#1 with GPT-5
Terminal-Bench 2.0
78.4%
vs. 64.7% Terminus 2
#1 with GPT-5.3-CodeX
SWE-Bench Pro
59.0%
vs. 40.2% SWE-Agent
#1 with Gemini-3-Flash
DevOps-Gym
46.8%
Only agent to complete end-to-end tasks
#1 DevOps

SageCTF

Built on OpenSage, SageCTF competed solo at DEF CON CTF 2026 — recovering 8 flags across 7 of the world's hardest challenges, placing in the top 5% of teams, and outscoring every No-AI and Low-AI team. Read the full story.

DEF CON CTF 2026
Top 5%
1,743 pts, 8 flags across 686 scored teams
50 CTF Benchmark
39/50
vs. 13/50 solved by Claude Code
Challenges solved
7/15
Solved 7 of 15 challenges attempted
gitvfs full solve
4.4%
Full-solve rate, only 30 of 686 teams