DensCor compresses agent-to-agent communication into dense 64-dimensional latent vectors — fast, precise, and fully auditable for regulated industries.
The Problem
Recent research proves that AI agents can communicate through compressed latent vectors — up to 24x faster than natural language. But these channels are opaque. No audit trail. No inspection. No compliance path.
The speed is real. The transparency is gone.
The coordination bottleneck — costs grow quadratically as agent count increases
Tokens per handoff in production agent systems
Latency per handoff via cloud LLM APIs
Visibility into what agents actually communicated
The Solution
DensCor.ai trains a Boundary Layer — a bidirectional translation module between natural language and a compact 64-dimensional latent vector. Agents communicate through this vector. Not through words.
64 numbers replace thousands of tokens. The Boundary Layer is trained with an explicit compactness objective — not as a byproduct of language modelling.
The latent channel outperforms natural language on accuracy. 98% vs. 78% on the same task. Less noise. More signal.
Every vector that crosses an agent boundary is logged — asynchronously, without slowing the chain. One inspection point for every compliance requirement.
Benchmarks
CLINC150, 500 samples, Qwen2.5-7B / 14B, A40 GPU.
| Metric | Natural Language | DensCor BL | Delta |
|---|---|---|---|
| Task accuracy | 78% | 98% | +20 PP |
| Latency per handoff | 0.5-2.5s | 30 ms | 28× faster |
| Data per handoff | 500-15k tokens | 64 floats | −99% |
| 3-agent chain accuracy | - | 73% | no text exchanged |
| Alignment cosine | −0.019 | +0.986 | +1.005 |
Verticals
Regulated environments need auditability at every step. The Boundary Layer makes this structurally enforced — not prompt-dependent.
Multi-agent clinical decision support with full audit trail. Compatible with Medical Device Regulation (MDR) and the EU AI Act.
Autonomous systems that coordinate agents must explain what was communicated, when, and why. Structurally enforced, not prompt-dependent.
Drop the Boundary Layer into any multi-agent workflow. Faster handoffs, lower inference costs, full audit trail.
Contact
Early-stage conversations with investors, partners, and enterprises building on multi-agent systems.