Agent harnesses tested for lifelong learning and skill growth
Continual-ARC: Measuring Skill Acquisition over a Working Lifetime
Machine Learning
Summary
Measuring how well AI systems learn new skills over time is difficult because most tests only check once. The authors created Continual-ARC, a collection of tasks that return over a long period, letting AI systems try again and learn continually. They tested different AI models and found that storing past examples often works better than trying to learn continuously inside the system without expert help. Also, learning new tasks well requires changes to the AI’s internal knowledge, which current harnesses don’t achieve.
What this means in practice
- •For ai system developers: Benchmark and compare continual learning methods on tasks that repeat over time with gaps and new challenges.
- •For machine learning engineers: Evaluate whether storing examples or updating model weights better improves performance for lifelong AI skills.
Authors
Daniel Lenton
Abstract
Agent harnesses have emerged as a promising mechanism for continual learning, alongside test-time parametric learning. However, the current benchmarks have several limitations: they generally focus on a single applied domain, serve each instance only once, and assume a single family of learner, so memory-based and weight-based systems are rarely compared on the same scale. We propose Continual-ARC, a lifelong stream built from the three ARC-AGI editions, whose tasks need no domain knowledge. The same tasks return after dormant gaps with fresh instances, while new tasks keep arriving. On each instance, the learner may pay for demonstrations, submit answers, or act inside the harness, with retries permitted on incorrect submissions, which are paid for. Each instance is scored based on the share of this feedback budget consumed. Since a learner may keep anything between instances, memory-based and weight-based systems are scored on the same scale. After the stream, the learner is frozen and tested on fresh instances of the tasks it met, which measures memorization, and on unseen tasks, which measures generalization. We run an agent harness with frozen weights, with and without persistence, on Qwen3.5-9B and GLM-5.3 Flash, comparing them with two prompting floors, a program-induction method and Qwen3.5-9B's weights trained to imitate GLM-5.3 Flash. Persistence in the harness generally does not improve memorization unless provided with explicit expert guidance. Most harness-based methods still learn less than a floor that simply caches each task's demonstrations, across all models we tested. No harness-based method gains on unseen tasks, whereas weight-based learning shows substantial gains. This opens the door to future weight-based continual learning that could teach models to drive harnesses in new domains.