01
Stage
CHANGES INITIAL STATEMoves the episode’s starting point to isolate a missing capability, add an obstacle, or remove already-mastered setup steps.
- Hide an object to force search
- Pre-complete early subgoals
- Calibrate task difficulty
Adaptive agent environments
Awakening static worlds for agent learning.
Keep the trusted environment frozen. Wrap its interface to expose the agent’s current weakness, validate the challenge with fresh rollouts, and turn the resulting experience into better skills or policy updates.
Maximum held-out gain+9.0ALFWorld OOD points
Execution reduction−9.8%SWE-bench average steps
SWE scaling gain+7.12At 300 environments
Supported wrappers3Stage · Contract · Chain
01 · Core thesis
Put a harness on both sides.
Agent systems already extend frozen models with tools, memory and skills. This paper applies the same external-layer principle to the environment side of the interaction.
A static benchmark keeps teaching what the policy already knows. An adaptive wrapper targets what it still cannot do.
THE ENVIRONMENT BECOMES A POLICY-CONDITIONED CURRICULUM.
Base system
Agent harnessFrozen language model
Environment harnessFrozen environment
External layer
Tools · Memory · SkillsAdd agent capabilities
States · Rules · CompositionReshape learning situations
Result
Autonomous agentModel + harness
Customized environmentStatic world + EnvHarness
Integrity
Model weights untouchedCapabilities remain modular
Verifier untouchedEvaluation stays grounded
02 · Programmable layer
Three ways to reshape a world.
Each component wraps the standard environment interface. They can be stacked, although their order matters because the transformations are non-commutative.
02
Contract
REWRITES INTERACTIONFilters actions and observations or changes transition rules to block shortcuts and enforce the intended behaviour.
- Mask verbose observations
- Require tests before submission
- Attach structured feedback
03
Chain
EXTENDS TASK HORIZONConnects multiple verified environments into a longer episode so the agent must preserve goals beyond a normal stopping point.
- Concatenate subtasks
- Train long-horizon persistence
- Combine component verifiers
03 · Under the scene
Watch EnvRigger build a curriculum.
The designer treats the policy as a black box. It diagnoses behaviour from rollouts, writes a targeted component, then accepts it only after fresh execution confirms a useful learning signal.
ENVIRONMENT DESIGN LOOP
Observe policy trajectories
FRAME 1 / 6AUTO PLAYING
04 · Experimental evidence
Targeting beats repetition.
Across embodied tasks, web interaction, software engineering and office automation, skills learned from customized environments consistently beat skills learned from the unchanged environments.
Held-out benchmark performance
EnvHarness versus skills extracted from the original environments. Bars show each benchmark’s reported primary metric.
EFFICIENCY SIGNAL55.01 → 49.61 steps
On SWE-bench Verified, the policy succeeds more often while using 5.40 fewer actions per episode. The interventions appear to repair loops and wasteful behaviour—not merely add task familiarity.
What improved
Absolute improvement over skills learned from unchanged environments.
| Benchmark | Metric | Δ |
|---|---|---|
| ALFWorld | Average SR | +5.9 |
| ALFWorld OOD | Success rate | +9.0 |
| WebArena | Average SR | +3.1 |
| SWE-verified | Resolved rate | +2.70 |
| OfficeQA | Exact match | +1.80 |
| SpreadsheetBench | Pass@1 | +3.27 |
Co-evolution keeps scaling
At the same 300-environment budget, adaptive batches continued improving while original and generated environments flattened.
EnvHarness · 54.79Original · 52.13
05 · Enterprise translation
From production failures to training worlds.
For regulated agent systems, the attractive pattern is to retain trusted controls and evaluation while making the learning scenarios responsive to observed failure modes.
Evidence challenge
Start the investigation with an incomplete bundle, add distractor documents, or pre-resolve routine identity fields so the agent must focus on material uncertainty.
Banking example: Hide the decisive shipping document among similar invoices to test evidence search rather than keyword matching.
Control enforcement
Reject an agent conclusion unless mandatory evidence, validation or screening steps have been completed and recorded in the trace.
Banking example: Block case closure until adverse-news evidence and source provenance are attached.
Workflow persistence
Connect document extraction, transaction investigation and report generation so the agent must preserve intent across handoffs.
Banking example: Require both evidence reconciliation and a defensible final recommendation before success.
06 · Critical read
Strong abstraction, bounded evidence.
The paper provides convincing evidence for adaptive wrappers, but it does not yet establish a universal environment-generation system for live enterprise workflows.
What is genuinely strong
01
Verifier preservationGenerated training signals remain grounded in trusted, human-built evaluation rather than an LLM-authored verifier.
02
Policy-conditioned curriculumEach new batch targets the equipped policy’s present boundary, explaining why performance continues to scale.
03
Cross-domain interfaceThe same reset/step abstraction works across five benchmarks and multiple model families.
04
More than skill promptingSeparate RL results show that reshaped environments can provide a useful direct optimization signal.
What remains unresolved
01
Resettable text worlds onlyIrreversible live actions, real accounts, visual GUIs and physical environments fall outside the current implementation.
02
Chain is not yet autonomousIt was excluded from the main EnvRigger pipeline and supports mainly sequential composition without semantic shared state.
03
Grounding is expensiveALFWorld consumed about 228M tokens versus 64.2M for GenEnv, although the latter relied on simulated rollouts.
04
Verifier ≠ realismA retained terminal verifier confirms success, but cannot guarantee every generated constraint is realistic or pedagogically ideal.
BOTTOM LINE
EnvHarness is best understood as adaptive environment middleware: a practical mechanism for keeping curricula near the agent’s moving capability boundary.
The deeper scaling equation is not only better models or larger skill banks. It is policy improvement + skill accumulation + environment co-evolution.