Modular systems have become the default architecture for many LLM agents operating on real services: order, payment, inventory, shipment. The problem is that observational traces show sequences, not authorizations. A world model trained on such data can learn that payment precedes shipment, but not whether payment actually authorizes shipment, whether inventory mediates the effect, or whether a hidden trigger explains both. FedCausalCompose enters precisely this gap: it is a causal world-model framework for modular agents in which local actions provide intervention-response evidence for cross-module interfaces.

The results define a stricter condition than it may appear. The authors show that observational world models accumulate an irreducible interventional error when unblocked back-door paths exist; interface recovery improves with intervention-response coverage; and an oracle causal composition can beat the non-causal lower bound only when coverage and local mechanism errors are controlled. In short, having a causal graph is not enough: the graph must be identified from intervention data rather than simple correlations, and local modules must be sufficiently reliable.

The most interesting part comes from the diagnostic experiments. In structured tool settings, where APIs expose preconditions and downstream effects, causal interfaces help clearly. In dialogue or narrative settings, agents tend to ignore raw edge lists unless a short attention anchor makes the causal information decision-relevant at action time. It is a result that moves the problem from statistical estimation to interface design: even a correctly identified causal structure is useless if the agent cannot use it when making a decision.

This is the thesis the research forces, and it should be read against the industrialization of agents. Companies are composing modular systems over internal APIs, often self-hosted or on-premise precisely to keep control of order, payment and logistics data. In this scenario, interfaces between modules are no longer simple REST contracts: they become causal contracts. The framework suggests that teams should collect not only observational logs but intervention evidence, for example what happens when an action is forced in one module and effects are observed in others. APIs that declare preconditions and effects become a prerequisite for reliable planning, not a documentation ornament.

There is also a second-order effect. If the value of a causal world model depends on how the information is presented at action time, then the bottleneck shifts from the model to the decision context. Agent frameworks may need to manage causal structures as active context rather than offline artifacts. Those designing orchestration pipelines will have to decide whether and how to inject causal constraints into the agent's prompt or memory; this carries TCO and maintenance costs because a causal knowledge base must stay aligned with real services. For those evaluating on-premise deployment of such systems, the lesson is twofold: infrastructure control helps generate and retain intervention data, but it does not remove the problem of agent-side readability. AI-RADAR covers some of these trade-offs at /llm-onpremise.