Oshan Ranasinghe

Writing / Designing Semantic Negotiation for Heterogeneous Intelligent Agents

System Breakdown

Designing Semantic Negotiation for Heterogeneous Intelligent Agents

April 5, 2026 · 8 min read

Why STRATON-LLM treats ontology mismatch as a negotiation problem rather than a simple alignment problem, and how the end-to-end system is structured.

STRATON-LLM started from a simple observation: in a multi-agent system, message delivery is not the same thing as shared meaning.

Two agents can exchange a perfectly valid message and still fail to cooperate because they interpret key concepts differently. One agent may ask for a price, while another internally reasons about a rate. The transport layer works. The semantics do not.

The real problem

In many systems, ontology mismatch is treated as a lookup issue:

  1. try direct term match
  2. try synonym match
  3. try ontology alignment model
  4. accept the best candidate

That pipeline is useful, but it is not enough for real interaction. A high-scoring match can still be wrong in context. A low-scoring match can still be correct if the agents exchange enough evidence. This is why STRATON-LLM treats mismatch handling as a negotiation problem.

The core idea

Instead of asking only “What is the closest term?”, STRATON-LLM asks a harder question:

Can two agents reach a trustworthy shared interpretation through a bounded, inspectable dialogue?

That shift changes the architecture.

From mismatch to negotiation

The system follows a six-layer flow:

| Step | Responsibility | | --- | --- | | Trigger & context | Detect the mismatch and gather local evidence | | Strategy selection | Choose the appropriate negotiation mode | | Proposal generation | Build claims, supports, attacks, and evidence requests | | Protocol engine | Control the finite-state dialogue | | Agreement evaluation | Decide whether the outcome is trustworthy | | Persistence & learning | Store mappings and traces for reuse |

This is the key design move in the project: semantic failure is elevated from a local parsing issue to a first-class system event.

Why not rely on raw LLM conversation?

Because raw conversation is flexible, but hard to trust.

An LLM can generate a plausible explanation for why price and rate are related, but the system still needs answers to these engineering questions:

  • Was the claim challenged?
  • Was evidence requested?
  • Did confidence improve consistently?
  • Did the agreement happen too quickly?
  • Can the session be replayed later?

Those questions cannot be answered reliably if negotiation is just a loose chat transcript.

The architecture principle behind STRATON-LLM

The project keeps LLMs inside bounded responsibilities.

  • LLMs assist with ontology matching support.
  • LLMs assist with proposal generation.
  • LLMs assist with argument reasoning.
  • But the protocol engine owns turn legality and state transitions.
  • The evaluation layer owns trust scoring.
  • The persistence layer owns mapping storage and learning.

That separation is what makes the system engineering-oriented rather than prompt-oriented.

A simple scenario

Imagine this interaction:

TravelAgent -> FinanceAgent: What is the price for premium package booking?

FinanceAgent does not use price as its native ontology concept. Its model is organized around rate.

A basic system might fail here, or silently guess.

STRATON-LLM takes a different path:

  1. check local ontology and mapping store
  2. call the matcher if confidence is low
  3. classify the mismatch
  4. launch negotiation if certainty is insufficient
  5. exchange structured argumentative acts
  6. score the final agreement
  7. persist the mapping if accepted

The important point is that resolution is earned, not assumed.

Why this matters beyond one project

This design matters because agent systems are moving toward:

  • cross-domain collaboration
  • independent agent development
  • dynamic capability discovery
  • decentralized coordination

In that world, agents will not always share one clean ontology. If semantic disagreement is ignored, the system looks intelligent only on happy-path demos.

What STRATON-LLM contributes

The project contributes a practical position:

  1. ontology matching is necessary but incomplete
  2. negotiation should be explicit and protocol-bound
  3. agreement quality must be evaluated after dialogue
  4. accepted outcomes should become reusable system knowledge

That is the foundation of STRATON-LLM.

The goal is not to make agents talk more. The goal is to make their agreements more trustworthy, replayable, and reusable.