If an AI keeps failing at a task, should you rebuild its practice world or reshape the one you trust?

This explores whether you should take an existing, trusted environment and reshape it around where an agent is struggling, or build a new one from scratch. The corpus has one direct example of the reshaping side and nothing that tests it head-to-head against building from scratch. EnvHarness wraps a static environment in a plug-in layer that sits on its reset and step calls. The layer can isolate individual skills, stretch task horizons and calibrate difficulty, all without touching the underlying code. Across five benchmarks it gained up to 9.0 points with 9.8% fewer steps (Can wrapping environments reshape how agents learn without breaking verifiers?). The main advantage is that the original human-built verifiers stay valid. In a new environment the verifier, which decides what counts as success, is usually the hardest part to get right, and you'd have to build and re-trust it.

The same pattern shows up elsewhere in the corpus: change what surrounds the fixed thing and leave the thing alone. Reorganizing a code harness around runtime behavior, rather than rewriting it, raised win rates by 10–19 points and let a weaker planner match stronger models at finding the right code (Can explicit behavior maps help weaker planners compete with stronger models?). Surveys of self-improving agents say recent progress is concentrated in these fast loops, which update prompts, memory and tools rather than model weights, because scaffold changes are cheaper and reversible (Do self-improving agents really split into two distinct loops?). A wrapper has the same property: if the reshaping doesn't help, you remove it. A from-scratch environment is a sunk cost. ModularRSI adds a related point. Harness improvements evolved as separate modules on benchmark-disjoint data carried over to unseen tasks and domains (Can harness modules improve separately from benchmark data?). By analogy, a reusable reshaping layer may travel further than an environment built for one task.

The "diagnose" half matters as much as the "reshape" half. An autoresearch pipeline gained 411% F1 on a memory benchmark, and its bug fixes, architectural changes and prompt rewrites each beat all the hyperparameter tuning combined. The authors' explanation is that it could read the code and reason about how the parts interact, which blind tuning cannot do (Can autonomous research pipelines discover AI architectures that AutoML cannot?). Skill creation shows a similar gap. Skills written inside the agent's running loop, with real feedback, beat skills authored offline, because the offline author never sees the situation where the skill fails (Does creating skills inside the agent loop eliminate mismatches?). Building an environment from scratch is the offline-authoring case, since you guess where agents will struggle before watching them struggle.

The loop has limits. Repeated revision tends to accumulate noise rather than improve, and the fix in one study was compressing memory between rounds (Do iterative refinement methods suffer from overthinking?). The Darwin Gödel Machine handles this by testing every change empirically and keeping an archive of variants instead of trusting one path (Can AI systems improve themselves through trial and error?). A reshaping loop needs the same discipline: measure each change and keep the alternatives. The corpus doesn't say when from-scratch is the right call, such as when no trustworthy verifier exists or when the base environment can't express the skill you want to train. It only shows that where a good verifier already exists, reshaping around it is the cheaper and safer bet.