Research map / Multi-agent systems and games / LLM agents

OpenSage: Self-programming Agent Generation Engine

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces OpenSage, an agent development kit where LLMs themselves create sub-agent topologies, write tools, and manage graph-based hierarchical memory, evaluated on CyberGym, Terminal-Bench 2.0, SWE-Bench Pro, DevOps-Gym, and LOCOMO.

Problem

Existing agent development kits require humans to manually design agent topology, toolsets, and memory, limiting generalizability and performance of constructed agents.

Approach

An ADK in which the backbone LLM autonomously creates and manages sub-agents at runtime (vertical and horizontal topologies with agent ensemble), synthesizes its own tools in isolated Docker sandboxes, and uses a hierarchical graph-based short- and long-term memory managed by a memory agent.

Key results

• OpenSage agent (SageAgent, GPT-5 medium) reaches 60.2% resolved on CyberGym, ranking first on the leaderboard and above Claude Opus 4.5 agent (50.6%).
• SageAgent with GPT-5.3-Codex achieves 78.4% ± 2.2 on Terminal-Bench 2.0, ahead of Droid (77.3%) and Simple Codex (75.1%).
• On SWE-Bench Pro Python subset (266 tasks) with Gemini 3 Flash, OpenSage reaches 59.0% vs. SWE-agent 40.2% and Agentless 9.4%.

Surprising findings

• Sub-agent creation with mismatched toolsets, hallucinated tools and sub-agents, and overly complicated instructions sometimes reduce effectiveness.
• Mem0g brings little improvement over no-memory on SWE-Bench Pro, which the authors attribute to hard-coded relationships and ungrounded node types.

Limitations

• Current SOTA models do not consistently use advanced features and sometimes misuse them.
• Ablations on CyberGym are run on a 300-instance subset rather than the full benchmark.
• LOCOMO results show OpenSage is not clearly superior to Mem0 variants (e.g., lower on multi-hop and single-hop).

Experimental setup

Datasets: CyberGym Terminal-Bench 2.0 SWE-Bench Pro DevOps-Gym LOCOMO

Benchmarks: CyberGym Terminal-Bench 2.0 SWE-Bench Pro DevOps-Gym LOCOMO

Metrics: % resolved accuracy (LLM-as-judge)

Baselines: OpenHands Claude Code SWE-agent Agentless Droid Simple Codex Mem0 Mem0g A-Mem Zep OpenAI memory

Builds on: Agentless Mem0 LOCOMO

Closest papers by meaning

• OpenHands: An Open Platform for AI Software Developers as Generalist Agents (ICLR 2025 · similarity 0.87)
• OpenAgentSafety: A Comprehensive Framework For Evaluating Real-World AI Agent Safety (ICLR 2026 · similarity 0.86)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.85)
• daVinci-Dev: Agent-native Mid-training for Software Engineering (ICML 2026 · similarity 0.85)
• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.85)
• Automated Design of Agentic Systems (ICLR 2025 · similarity 0.84)