Cross-Workflow Latent Communication for Heterogeneous Agents

We study whether one set of latent communication interfaces can support heterogeneous agents in different agent workflows. Both workflows, all six model assignments, and a shared instruction pool are included in training and evaluation on disjoint tasks. We introduce (Reusable Cross-Agent State Transfer). A Latent Transition Adapter generates continuous inputs within each agent, and a directed Cross-Model Bridge transfers every generated intermediate-layer state to the next agent. Training combines multi-role warm-up, joint task and intermediate-readout supervision, and continuous-action group-relative policy optimization. Gaussian latent exploration is confined to reinforcement learning; deployment follows the deterministic mean path. Across six benchmarks, the reported average scores are 33.9% and 34.6% for the two workflows, exceeding the strongest aggregate comparator by 2.0 and 2.1 percentage points. Removing warm-up reduces scores by 3.4 and 3.8 points; disabling communication reduces them by 3.9 and 4.2 points. These comparisons support shared latent interfaces within the trained configuration support.