$ cat wiki/papers/2026/2608.20634-agent-mercury.md
AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale (arXiv:2608.20634)
TL;DR
Instead of building an environment for a task, AgentMercury instantiates a persistent world — entities, services, tools, state and executable cross-service invariants — from a high-level business scenario, and lets tasks emerge from it. 4,783 executable environments across 14 industries and 50 countries were built and used as RL training substrate. Two results carry the paper: training on them transfers out of domain, and the construction process itself is learnable — authoring success on held-out scenarios rises 3.3% → 83.3% after fine-tuning on construction traces (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED from this run's sandbox and
the paper was not read. The experiments use Qwen3.5-4B and
Qwen3.5-35B-A3B, which indicates a lab working on the Qwen line but does
not establish Alibaba / Qwen AI Lab authorship — open weights are open to
everyone
(source).
Listed on HuggingFace Daily Papers, 2026-08-26, 9 upvotes — that community's popularity signal and nothing more (source).
Method
The gap it names. Training environments are "manually constructed or synthesized around predefined tasks and benchmarks". That task-centric paradigm cannot scale to realistic evolving workflows, because the workflow is what generates the tasks, not the other way round.
The inversion. AgentMercury first instantiates a persistent world holding:
- entities
- services
- tools
- state
- executable cross-service invariants
Diverse tasks and interaction trajectories then emerge from that world.
Scale built: 4,783 executable environments, 14 industries, 50 countries.
The second experiment treats environment construction as a task in its own right: fine-tune on construction traces and measure whether the model can author a working world for a held-out business scenario.
Results
| Setting | Before | After |
|---|---|---|
| Qwen3.5-4B on EnterpriseOps-GYM | 12.3 | 15.7 |
| Qwen3.5-4B on AIME26 | 45.9 | 56.0 |
| Qwen3.5-35B-A3B, executable-world authoring on held-out scenarios | 3.3% | 83.3% |
| The AIME26 row is the one that does the arguing: **+10.1 points on competition | ||
| mathematics** from training on business environments the benchmark had nothing | ||
| to do with. The paper states the environments were "generated **without targeting | ||
| the evaluation benchmarks**", and reports gains across enterprise workflows plus | ||
| out-of-domain reasoning, coding, scientific computing and tool use. |
Significance
It is Environment Scaling with a number attached, on the same day the idea is asserted without one. Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283) names Environment Scaling as one of its two axes and reports no figure; AgentMercury reports 4,783 environments and a measured out-of-domain transfer. Two papers in one snapshot, one arguing the thesis and one testing it.
The 3.3% → 83.3% row is the more consequential half. If constructing verified environments is itself learnable, the bottleneck on Agentic Reinforcement Learning moves from human-authored environments to compute, which is a different kind of constraint. That is the claim to watch, and it rests on one model and one held-out set.
+10.1 on AIME26 from business workflows needs a mechanism and does not have one. The obvious readings — that the environments teach general multi-step verification, or that the 4B baseline was undertrained on any long-horizon data — are not separated by anything in the abstract, and the second would make the result much less interesting.
It joins the cluster Agents (LLM Agents) has been accumulating for two weeks: the scaffold, the environment and the harness are being scaled as first-class objects while the weights stay where they are.
Open Questions
- What generalisation is actually being measured? Environments generated "without targeting the benchmarks" is a claim about intent, not about distribution overlap, and no contamination analysis is mentioned.
- Who verifies the invariants? "Executable cross-service invariants" is the load-bearing verifiability claim; how they are checked, and whether a synthesised invariant can be wrong, is unstated.
- 83.3% of what? Authoring success on held-out scenarios needs a definition of success — does the world merely run, or does it hold its invariants?
- Does it hold on a frontier model? The gains are reported on a 4B and a 35B-A3B. Whether an environment corpus this cheap moves a frontier model at all is the question that decides how much this matters.
- Cost. 4,783 environments is a compute figure nobody gives.
Cite
AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business
Scenarios at scale (2026). arXiv:2608.20634.