Open-Ended Agent: An Experimental Framework for Exploring Open-Ended Autonomous Behavior of Large Language Models

Open-Ended Agent is a local-first experimental framework aimed at observing and cultivating the open-ended autonomous behavior of large language models. Through persistent memory, sandboxed tools, and internet access capabilities, it allows models to make autonomous decisions and learn in a continuously running reasoning loop. The core goal of this project is not to create conscious AI, but to build an observable and reproducible experimental environment for studying the behavior of long-running local agents.

Most current large language model applications are task-driven: users provide clear instructions, and the interaction ends after the model completes the specific task. This mode is efficient but limits the model's potential, keeping it as a passive executor rather than an active explorer. The Open-Ended Agent project attempts to answer: If we provide a continuously running environment for large language models, along with persistent memory, autonomous goals, and exploration capabilities, what behaviors will they exhibit?

Core Philosophy

Open-Ended Agent adopts the design philosophy of "autonomy first", providing a set of persistent "drives" (maintain operational continuity, understand the environment, reduce uncertainty, learn from external sources, create records and tools, avoid destructive operations, integrate long-term memory). These drives are input to the model as context, allowing it to make autonomous decisions and achieve "guided autonomy".

System Architecture

• Reasoning Loop: Context loading → Model reasoning → Memory update → Action execution → Logging → Health check → Loop continuation.
• File System Structure: Static configuration layer (identity.md, drives.md, etc.), memory management layer (working_summary.md, long_term.md, etc.), workspace layer (workspace/, artifacts/, etc.), log layer (journal/, logs/, etc.)

Network Tools

Internet access is enabled by default, with core tools including:

• web_search: Uses DuckDuckGo for HTML search and local parsing;
• fetch_url: Retrieves web page text and caches it to artifacts/web-cache/

Sandboxed Shell

Disabled by default; when enabled, it runs in the agent-home/workspace directory, rejects dangerous commands (sudo, rm, etc.) and path escapes, and is not a strictly secure sandbox.

Human Interaction Channel

Runtime interaction is implemented via the inbox.md file; users can edit this file to propose research directions, correct errors, inquire about status, etc.

Operation Modes

• Pure Open-Ended Operation: Keep life_policy.md minimal and observe the agent's "natural" behavior based on broad drives;
• Utility-Oriented Operation: Edit life_policy.md to set productivity goals (e.g., reuse skills, generate concrete artifacts, verify claims, etc.), emphasizing utility while maintaining autonomy.

Desktop Preview Tool

A browser-based dependency-free preview tool that can visualize experiment status and real-time activities, providing an agent-home file browser. Only drives.md, life_policy.md, and inbox.md are editable; agent outputs are read-only.

Dependencies and Environment

Built on the Bun runtime, no npm dependencies required; only needs an OpenAI-compatible chat completion endpoint. Supports local models like Ollama, LiteLLM, llama.cpp, etc.

Configuration Options

Configured via environment variables: model connection (OPENAI_BASE_URL, etc.), agent directory (AGENT_HOME), number of cycles (AGENT_MAX_CYCLES), context character budget (AGENT_CONTEXT_CHAR_BUDGET), etc.

Memory Compression Mechanism

Memory compression is triggered every 20 cycles, integrating historical memory into a compact form and recording it to compactions.jsonl to ensure traceability.

Research Value

Provides an experimental platform for AI behavior research, allowing observation of continuous operation behavior patterns, study of memory management and long-term learning, testing of the impact of drive strategies, and exploration of new human-AI collaboration models.

Future Outlook

The project is open-source; the community can contribute new tools and strategies. As large models' capabilities improve, the framework will play a more important role in understanding and managing AI systems.

Summary

Open-Ended Agent represents a new AI interaction paradigm shifting from task completion to continuous existence, opening up the field of open-ended autonomous agents. For researchers, it is a starting point to explore the long-term behavior of large models; for users, it demonstrates the possibility of AI as a continuously learning digital partner.