The Boundary Layer addresses a gap that the research community has identified but not yet closed.
Architectural Lineage
| Paper | Finding | Limitation |
|---|---|---|
| LatentMAS Zou et al., 2025 |
Training-free latent collaboration: 83% token reduction, 4× speedup | Built on hidden states optimised for language, not communication. No audit trail. |
| Interlat Du et al., 2026 |
Learned compression of hidden states: 24× inference speedup | Ad-hoc layer selection; training objective is downstream task, not communication fidelity. |
| Vision Wormhole Liu et al., 2026 |
Cross-model communication via VLM interface | Requires VLM backbone; ~0.05B parameter codec per model pair. |
| Coconut Hao et al., 2024 |
Continuous latent reasoning in single-agent setting | Does not address inter-agent communication. |
| Multi-Agent Teams Pappu et al., 2026 |
Multi-agent teams can underperform single models due to coordination failures | Defines the coordination problem; does not solve it. |
Differentiation
Training objective is communication fidelity, not language modelling. The protocol is a first-class citizen.
Encode → Decode → Re-encode ≈ original. Agents can respond in latent space, not just receive.
Every agent boundary is logged. Asynchronous. Human-readable via k-NN lookup. Built in, not bolted on.
Two independently trained models communicate through the same protocol space.
Atomic facts travel as typed key-value pairs — lossless, never compressed.
Three endpoints. Pay per vector. No idle cost. Scale to zero.
Open Questions
At what compression does task performance degrade per domain? The 64-dimension choice is empirically derived. The optimal may vary.
Do protocol spaces align without explicit cross-model supervision when trained with the same objective? Early results are promising.
Do efficiency gains increase with model scale? Theory says yes. Empirical validation at 32B+ is pending.
What information is preserved and what is discarded? A mechanistic interpretability question with direct safety relevance.