AgentMercury

84-second overview: how a business scenario becomes an executable world, how policies train inside it, and what transfers. All curves and numbers are from the actual runs.

σ → Planet → w → Task → (u, ρ) → π → τ → Grade → r

A scenario σ is compiled by Planet into an executable world w; tasks (u, ρ) are instantiated from the world; a policy π produces a trajectory τ that is graded deterministically into a reward r.

Abstract

Policy models learn to select actions; world models learn to predict subsequent states. Both depend on diverse, informative environments for training — yet environments are typically hand-built or synthesized around predefined tasks and benchmarks. This task-centric paradigm limits environments to what the target benchmark measures. AgentMercury flips the direction: it synthesizes executable environments from high-level business scenarios, making environment construction an explicit Planet role in the agent–environment system. Policies trained with RL inside these worlds show steadily improving rewards and transfer to established benchmarks spanning knowledge, reasoning, and tool use — despite the environments being built independently of every evaluation task. The construction traces themselves are learnable: a fine-tuned open model authors new valid worlds through code-level modifications, suggesting a scalable loop of environment generation and agent learning.

4,783executable worlds

14industries · 50 countries

13.98Mseeded state rows

842unique tools

43,300verifiable RL tasks

AgentMercury spans 14 industries and 50 countries with 4,783 environments; each environment averages 13 services, 65 tools, 31 state tables, and 15 SQL verifiers. — **Diverse, multi-service, verifiable.** (a) 4,783 environments across 14 industries and 50 countries. (b) Against existing agent environments, AgentMercury occupies the regime that is both large-scale and multi-service. (c) A typical world carries ~13 services, ~65 tools, ~31 state tables, and ~15 deterministic SQL verifiers over one shared state database.

How Planet builds a world

Planet receives only a scenario σ and emits a complete world specification w = ⟨S, A, Ω, T, O, s₀, R⟩, factorized as grounded company identity → service graph → state schema → seeded initial state → world-level invariants. Crucially, the invariants R are not enforced by the transition function — they are rendered as executable verification conditions (largely SQL over the world’s own database). The agent is responsible for satisfying cross-service requirements; the world only checks them afterwards. One world then serves many tasks: task instantiation seeds task-specific state Δs₀ and a rubric ρ on top of the shared world, and post-episode grading is deterministic and replayable.

Pipeline from scenario to executable world to task instantiation, agent interaction, and deterministic post-episode verification. — **(A)** Outside the world: Planet constructs company identity, service graph, state schema, seeded state, and invariants; Task instantiates objectives. **(B)** Inside the world: the policy acts from history while the environment transitions deterministically; a post-episode verifier grades the trajectory with the task rubric and hidden invariant views.

Corpus composition: 43,300 tasks from 4,326 environments; the RL run touches 3,200 tasks yet retains over half of the structural diversity. — **Worlds are built once; tasks are many.** The RL corpus holds 43,300 investigation tasks from 4,326 company environments (2,287 industries, 842 tools, 222 state-table types, 148 service combinations). The reported 4B run receives gradients on only 3,200 tasks (7.4%) — yet still covers 53.5% of environments, 62.9% of industries, and 75.8% of tools.

Training inside synthesized worlds

We train Qwen3.5-4B and Qwen3.5-35B-A3B with RL (GRPO, and SAO as an alternative optimizer) directly in the synthesized environments — multi-turn MCP tool use against live business software, rewarded by the deterministic environment grader. Rewards improve steadily; the truncated-response ratio collapses; degenerate responses stay near zero. The improvement is progressive across checkpoints rather than a single-step jump.

Reward rises from 0.25 to about 0.53 for Qwen3.5-4B GRPO over 149 steps and from 0.35 to about 0.64 for Qwen3.5-35B-A3B SAO over 180 steps, while truncation falls. — **Healthy optimization at two scales.** Left: Qwen3.5-4B + GRPO — mean reward roughly doubles (~0.25 → ~0.53 at peak). Middle: Qwen3.5-35B-A3B on the same worlds — GRPO climbs 0.42 → 0.57 and SAO 0.34 → 0.64. Right: the truncated-response ratio collapses in all three runs (78% → 3% for the 4B) while degenerate responses stay near zero. Thin lines: raw per-step values; bold: smoothed. 4B curves are digitized from the paper’s training figure; 35B curves are parsed from the raw run logs.

Out-of-domain benchmark performance rises across training checkpoints on AIME26, HMMT, LiveCodeBench, SciCode, tau3, BFCL, and GPQA-Diamond. — **Transfer emerges during training, not at a lucky checkpoint.** Out-of-domain performance of Qwen3.5-4B + GRPO across checkpoints; dashed line = base model, error bars = s.d. over three independent evaluation runs.

Results

In-domain (EnterpriseOps-Gym, 8 enterprise domains). Training on AgentMercury improves the 4B average from 12.3 to 15.7 (+27.6%, 7 of 8 domains) and the 35B-A3B average from 24.8 to 28.1 (GRPO) / 28.3 (SAO) — with SAO improving all eight domains. None of the evaluation environments were seen during training.

Model	Teams	CSM	Email	ITSM	Calendar	HR	Drive	Hybrid	Avg.
GPT-5†	26.3	36.4	49.0	18.9	41.3	17.9	34.0	23.5	30.9
Gemini-2.5-Pro†	39.3	11.6	31.1	13.9	12.5	4.9	27.0	19.6	19.9
Kimi-K2-Thinking†	30.0	7.1	51.0	12.2	15.4	8.2	39.6	15.7	22.4
Qwen3.5-4B	20.8	9.2	23.9	6.8	10.4	9.5	6.2	11.7	12.3
+ GRPO + Ours	23.0	5.6	33.3	7.4	13.1	10.8	15.6	17.0	15.7
Qwen3.5-35B-A3B	33.9	7.6	53.2	14.6	18.6	11.4	35.9	23.5	24.8
+ GRPO + Ours	39.8	11.1	53.8	15.6	21.9	17.7	41.2	23.7	28.1
+ SAO + Ours	39.1	12.5	54.8	16.5	22.6	17.7	39.1	24.1	28.3

† reference numbers reported by EnterpriseOps-Gym. Ours: mean over three independent runs; s.d. in the paper.

Out-of-domain. The same policies improve on benchmarks the environments were never built for — math, coding, scientific computing, knowledge, and independent tool use:

Model	AIME26	HMMT	LCB v5–6	SciCode	τ³-Airline	τ³-Retail	τ³-Telecom	BFCL	GPQA-D
Qwen3.5-4B	45.9	28.5	36.6	22.6	48.8	70.4	92.5	30.3	76.5
+ GRPO + Ours	56.0 +10.1	35.4 +6.9	44.0 +7.4	25.7 +3.1	58.7 +9.9	73.6 +3.2	91.9 −0.6	31.7 +1.4	77.5 +1.0
Qwen3.5-35B-A3B	91.0	77.0	74.3	29.7	39.1	52.7	49.1	31.1	82.8
+ GRPO + Ours	91.9	80.0	79.0 +4.7	29.9	52.5 +13.4	57.1 +4.4	68.9 +19.8	42.5 +11.4	85.0 +2.2
+ SAO + Ours	92.2	83.3 +6.3	78.6	28.3	50.9 +11.8	55.7	65.5 +16.4	42.1 +11.0	84.0

Mean over three independent evaluation runs (± s.d. in the paper). Beyond the means, the 35B policy becomes far more consistent: τ³-Telecom run-to-run s.d. shrinks from ±49.8 to ±20.5 after training.

Agents can learn to author worlds

Because AgentMercury exposes its construction traces — brief-to-world generation, intermediate-stage completion, validator-guided corruption repair, and intent-to-diff prediction — environment authoring itself becomes a learnable task with an executable oracle: a generated world counts only if it passes all 12 structural validators.

Fine-tuned on 29,823 construction-trace samples, Qwen3.5-35B-A3B goes from passing the full authoring oracle on 3.3% of held-out business briefs to 83.3% — level with the strongest API models (Claude Opus 4.8: 83.3%, DeepSeek-V4-Pro: 83.3%, GLM-5.2: 80.0%, GPT-5.4: 66.7% zero-shot); Fisher exact p = 1.2×10⁻¹⁰. The dominant remaining failure everywhere is cross-service structure — exactly what the executable oracle exists to catch.

Prompting and learning turn out to be different levers: an in-context construction recipe helps the raw base model (3.3% → 20.0%) but collapses the fine-tuned one (83.3% → 10.0%) — once the procedure is internalized in the weights, the recipe interferes with it. Worlds train agents; trained agents can build new worlds. That closes a scalable loop between environment generation and agent learning.

Release

Everything is distributed through one place — the AgentMercury collection on Hugging Face:

Trained policy models — the RL-trained checkpoints
Environments & tasks — business scenarios, task instructions, and the RL corpus
Construction code-diffs — the code-level differences that instantiate each environment

BibTeX

@article{jeong2026agentmercury,
  title   = {AgentMercury: Your Agent Can Synthesize Verifiable Environments
             for Business Scenarios at Scale},
  author  = {Jeong, Minbyul and Yoon, Chanwoong},
  journal = {arXiv preprint},
  year    = {2026}
}

Last updated 2026-08-19.