CompoWorld: Compositional Environment Scaling Boosts Agent Performance by 9.17 Points

Cross-App Tasks Demand Compositional Training Environments

Real-world tasks often require an agent to chain multiple applications: query a database, read a manual, update a record, then notify stakeholders. Each step depends on information produced by the previous one. Existing work builds isolated tool environments, but simply adding more tools to a flat list does not guarantee meaningful cross-service information flow. CompoWorld addresses this by treating each service as an independent, reusable component that can be wired together via explicit information dependencies.

Building 448 Verified, Reusable Services

Starting from the public MCP tool specification, the team collected tool names, descriptions, and parameter types. A Coding Agent then generated a mock service for each, encapsulating its own state, tool set, and state-transition logic. Services share a unified Python type model and a consistent JSON execution convention, enabling them to run together after composition. Each service undergoes a two-stage verification: the Coding Agent writes implementation and tests in isolation; then an independent Agent session creates additional tests against the tool spec to check edge cases. Only services passing both stages are marked as deterministic. For tools that rely on external systems (7.1% of 10,130 tools), a world-model LLM simulates returns and state updates, which still pass the same type checks. The final library comprises 448 services covering 10,130 tools across diverse domains.

Generating Tasks via Service-Level Dependency Graphs

CompoWorld uses a random-walk process to select and connect services into a service-level dependency graph . Nodes are services; edges denote that the downstream service requires information from the upstream service's output. This graph captures realistic cross-app data flow (e.g., a database query result determines which document to read, which in turn dictates the notification content). Tasks may revisit a service after state changes, forcing the agent to re-query and synchronize information. Agents discover services and tools at runtime via meta-tools ( list_services, list_tools, describe_tool, call_tool). Task generation proceeds in four phases:

Explore services – actually call tools to learn interfaces and state structures.

Design task – craft initial state, user instruction, and acceptance criteria based on the dependency graph.

Execute verification – run the task with real tool calls to confirm achievability and that the evaluator accepts the reference result.

Submit task – store the verified instruction, environment snapshot, and evaluation artifacts.

With 448 services, choosing 5 at a time yields ~1.47×10¹¹ candidate combinations; validity is determined by the generation-and-verification pipeline.

Completion-Focused Rubric Reward for Reinforcement Learning

Cross-app tasks have multiple acceptance conditions. A binary success/fail reward obscures partial progress; uniform per-condition rewards over-weight easy conditions. CompoWorld introduces Completion-Focused Rubric Reward : within each sampling group, conditions with lower pass rates receive higher weights. For example, if most trajectories update records but few send notifications, the notification condition gets a larger weight. The model still earns reward for partial completion, but is pushed toward the currently weakest requirements. This reward is combined with GRPO, with weights recomputed per sampling group.

Experimental Results

Base model: Qwen3.6-35B-A3B . Training data: 3K SFT trajectories and 1K RL tasks from CompoWorld environments. Evaluated on eight benchmarks covering business processes, planning, tool use, and long-horizon tasks.

Average improvement over base: +9.17 points .

On AutomationBench (cross-app business workflows): success rate 10.33% → 32.33% (+22.00 pp), surpassing GPT-5.4 (27.67%), Claude Opus 4.6 (25.50%), and the best 35B-A3B specialist model Occamy-1.0 by 4.73 pp.

Ablation: Compositional vs. Single-Environment Training

Under identical SFT pipeline, environment count, and 1K trajectory budget, compositional training outperformed single-environment training on all four completed comparisons. AutomationBench: 12.83% → 33.33% (+20.50 pp). τ³-Banking: 9.62 → 17.53 (+7.91). This confirms that cross-service dependencies in the training data provide additional gains.

Scaling Analysis: More Compositions ≠ Monotonic Gains

SFT-only experiments with 100, 500, 1,000, and 3,000 training samples (each a distinct compositional environment) showed all four benchmarks beat the base at 100 samples, but further scaling did not consistently improve every benchmark. This indicates that scaling composition count must be paired with service quality, task selection, and verification rigor.

Conclusion and Outlook

CompoWorld provides a scalable path for general agents to master cross-application interaction: build services as reusable components, generate tasks via dependency graphs, verify executability, and train with a reward that focuses on unfinished requirements. As the service library grows, each component can participate in more combinations, and existing components remain reusable. Future work includes selecting higher-value compositions and verifying more complex information dependencies.

Paper: CompoWorld: Compositional Environment Scaling for General Agents (arXiv:2609.33665). Code: https://github.com/AllSpark-Research/AgentEnv.