From Long-horizon Tasks to Endless Frontier Tasks. The next frontier for LLM agents may not be merely completing longer and more complex workflows, but continuously improving open-ended objectives with no fixed endpoint. These tasks require agents to autonomously conduct broad research, propose hypotheses, design and run experiments, judge their own results, identify failure modes, refine strategies, and repeat this loop again and again. The goal is not to finish a task, but to keep pushing the frontier. Examples include exploring mathematical theories, discovering scientific hypotheses, finding market alpha, improving model training recipes, optimizing codebases, designing better products, running autonomous security research, and even improving agent systems themselves. Long-horizon tasks test whether agents can complete a difficult workflow. Endless Frontier Tasks test whether agents can sustain discovery, optimization, and self-improvement over time.

Strongly agree that true usefulness matters more than benchmark points. That’s exactly why it is encouraging to see GLM-5.2 improve in real development scenarios, not just on public evals. The interesting contrast is not benchmark vs usefulness, but models that keep moving on both and models that have been oddly quiet on both for a long time. More here:

GLM-5.2 is not only stronger on benchmarks, but also much better in real app development scenarios — iOS, Android, WeChat Mini Programs, and more. Behind this jump is a full loop from environment construction, evaluation, data optimization, reward design, to training. Real tasks, real execution, real improvement.