In this article, we introduce a Reinforcement Learning (RL)–driven framework for Information Bottleneck (IB)–based distributed Joint Source–Channel Coding (JSCC) that operates reliably when the forward channels are unknown, non-differentiable, stochastic, or entirely black-box. Current state-of-the-art deep variational IB methods require differentiable end-to-end models and full knowledge of the forward channels’ statistics, which renders them ineffective in many practical scenarios involving hidden channel states, human-in-the-loop setups, or proprietary simulators. To overcome these limitations, we reformulate the IB-based JSCC design problem as a sequential decision-making task. This enables the use of deep Multi-Agent Reinforcement Learning (MARL). We first revisit the single-terminal setup and show that the encoder can be treated as a policy in a contextual bandit whose sampled reward is set as the decoder reconstruction term, while the compression penalty is applied directly as an analytic actor regularizer derived from a variational surrogate of the IB objective. This yields a model-free compressor that learns to preserve relevance while respecting the rate constraint, without requiring gradients through the channel. We then extend the framework to the multiterminal setting, where multiple encoders observe noisy versions of a common source and must coordinate implicitly through the environment. Two retrieval strategies are considered: a parallel scheme, which ignores the side-information at the decoder, and a successive scheme, which exploits the side-information at the decoder via conditional priors. For both scenarios, we derive RL-compatible variational lower-bounds on the original IB objectives, enabling Centralized Training with Decentralized Execution (CTDE). By this, we generalize state-of-the-art distributed data-driven IB-based JSCC schemes to arbitrary forward channels while retaining the scalability and sample efficiency. As the main highlight, this work demonstrates that MARL provides a principled foundation for learning distributed compressors in environments where model- or gradient-based approaches are fundamentally inapplicable.