How It Works — DensCor.ai

The Bottleneck

AI agents are fast. The space between them isn't.

When an AI system completes a complex task, it rarely uses a single model. It uses a chain: one agent breaks the task down, another executes a subtask, another verifies the result.

Every connection in that chain is a bottleneck. Each agent must receive, parse, generate, and pass full natural-language text. This is O(N²) overhead. The Boundary Layer makes each connection 28× cheaper.

// Traditional: O(N²) text overhead
Agent A → 847ms → Agent B
Agent B → 847ms → Agent C
Agent A → 847ms → Agent C
// DensCor BL: compact vectors
Agent A → 30ms → Agent B
Agent B → 30ms → Agent C
Agent A → 30ms → Agent C

The Protocol

What a vector carries.

The Boundary Layer compresses the full contextualised meaning of a message into a 64-dimensional vector. This is not a summary. It's a geometric representation of intent — trained with three simultaneous objectives.

Reconstructability

The receiving agent must be able to execute the correct task from the vector alone. No additional context needed.

Task Fidelity

The vector must preserve enough information that the downstream agent reaches the same outcome as it would from the full text.

Compactness

The vector dimension is minimized subject to the first two constraints. The model learns to use as few dimensions as the task allows.

Slots: Atomic facts — names, numbers, dates, identifiers — travel as typed key-value pairs alongside the vector, never through compression. Lossless over the entire agent chain.

Three Modes

Three modes. One protocol.

Mode A — Encode

Text → [BL vector + Slots]

Natural language in. Compact vector out. The Transformer runs once. Structured facts travel as typed slots — never compressed.

Mode B — Execute

[BL vector] → [BL vector]

Agent to agent. No Transformer. No tokens. The Boundary Layer decoder maps the incoming vector directly to the next task. This is the fast path.

Mode C — Render

[BL vector] → Text

Latent back to language. Happens once, at the end of the chain, for human-facing output.

Auditability

Every handoff. Inspectable. Asynchronous.

At every point where a vector crosses an agent boundary, a log entry is written. The log translates the vector back to human-readable language via k-NN lookup.

The log runs asynchronously — it does not slow the chain. It is the structural basis for EU AI Act compliance, MDR auditability, and enterprise security requirements.

[14:22:01] Agent A → Agent B

              intent: book_hotel  
              confidence: 0.94

              slots: { hotel: "Standlerhof", checkin: "03-08" }
            
[14:22:01] Agent B → Agent C

              intent: availability_confirmed  
              confidence: 0.91

              slots: { available: true, price: 180 }

Efficiency

Architectural efficiency.

28×

Speedup

Latency between handoffs: 847ms → 30ms per connection

−99%

Data reduction

64 floats instead of ~74 tokens per handoff

0.986

Alignment cosine

Encode → Decode → Re-encode fidelity

Engineering the DensCor.ai core.