What's this about?
Two papers on the same bottleneck. SPADE (42 upvotes) argues continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals, while existing training environment pools are hand-curated, statically synthesized, or use frozen verifiers. EnvHarness (22 upvotes) puts the same point more vividly: agents learn by interacting with environments, yet those environments are hand-built and static — blind to an agent's weaknesses, and quickly left behind as it improves.
What exactly is a "training environment" for an agent?
It's a simulated world with tasks, tools and a way to score success — a fake filesystem to navigate, a mock API to call, a codebase with planted bugs. Environments are to agent training what datasets are to model training, and they have the same problem in a worse form: a dataset is at least fixed while the agent improves past it, whereas an environment is supposed to keep challenging the agent and can't, because nobody rewrites it. So an agent saturates its environment and stops learning while looking like it's still training.
Why is this worth noticing?
- "Blind to an agent's weaknesses" is the precise complaint. A static environment cannot notice what this particular agent is bad at, so training effort goes wherever the environment happens to point.
- It's the missing half of every self-improvement result this month. DarwinX, StateM and Zetta all evolve the agent or its harness; these ask who evolves the curriculum.
- The frozen-verifier problem is the subtle one, since an environment whose grader can't keep up will reward gaming rather than capability — the same lesson as Every Model Cheats, in the training loop rather than the benchmark.