About Anchor Research
A non-profit research effort focused on understanding and improving the long-term behavior of autonomous AI agents.
Our Mission
Modern AI agents can plan, make decisions, and interact with the world in complex ways. They are increasingly being tested in tasks that last many hours or days, sometimes operating with minimal human supervision.
Current safety checks often measure short-term or one-shot tasks. But if an agent runs continuously for a week, will it still follow its intended instructions on day seven as it did on day one?
Without long-term testing, we risk discovering that an AI agent seems fine in short bursts, only to develop problematic behaviors when deployed for extended timelines.
Key Focus Areas
- Behavior Drift: The agent gradually shifting away from its original goals.
- Unsafe Failures: Getting stuck in cycles and failing to progress safely.
- Risky Adaptations: Adopting new strategies that might be unsafe or misaligned.
Our Approach
We combine rigorous evaluation frameworks with open-source tools and collaborative research to advance long-term AI safety.
Evaluation Frameworks
We design ways to observe agent behavior across different time scales, creating test environments that gradually change to see how agents adapt.
Open-Source Tools
Building prototypes like our long_agent_framework to help others run extended tests, gathering data on agent decisions and resource use.
Collaboration
Working with AI labs and research groups for deeper insights, openly sharing findings and best practices with the community.
Our Work
Our primary output is the open-source long_agent_framework, a Python-based toolkit for evaluating the behavior of AI agents over extended periods. We also maintain a collection of research and resources at our GitHub organization.
Our Goals
Identify & Document Long-Horizon Issues
Proactively discover where AI agents diverge from intended goals or form hazardous strategies over time.
Develop Standardized Metrics
Create clear ways to measure how stable and adaptable an agent is across different time scales, enabling comparison of results across labs.
Establish Long-Term Evaluation as a Norm
Encourage AI developers to see multi-day or multi-week testing as a standard part of deploying advanced agents.
Inform Real-World Safety Efforts
Provide evidence-based insights that help both researchers and policymakers understand the risks of truly autonomous AI systems.
Team
Diogo Cruz
Founder & Lead Researcher
Background
- • PhD-level background in quantum computing
- • Subsequent pivot to AI safety research
- • Technical AI Safety Researcher
- • Experience in multiple AI safety projects
Experience
- • Led research teams on neural network analysis
- • Multi-turn jailbreak evaluations
- • Safety-focused tool development
- • CHAI Summer Internship
Why We're Different
Exclusive Focus on Long-Run Behavior
Unlike many projects that look at short tasks, we zero in on extended, multi-day or multi-week runs.
Open & Collaborative
As a non-profit effort, our goal is to share tools and results widely, fostering a safer AI ecosystem.
Technical + Research Partnerships
We aim to keep our work grounded in practical concerns through engagements with the broader research community.
Ready to Collaborate?
Join us in making AI systems safer through rigorous long-term evaluation.
Get Started