Oshan Ranasinghe

Writing / Why Ontology Alignment Alone Is Not Enough for Real Agent Interoperability

Concept to Implementation

Why Ontology Alignment Alone Is Not Enough for Real Agent Interoperability

April 4, 2026 · 7 min read

Ontology alignment helps agents find candidate correspondences, but real interoperability needs context, confidence handling, negotiation, and post-dialogue evaluation.

Ontology alignment is often presented as the answer to heterogeneous agent communication.

It is an important part of the answer, but not the full answer.

What ontology alignment does well

Ontology alignment is good at producing candidate correspondences such as:

  • price -> rate
  • customer -> client
  • booking -> reservation

That is valuable. It reduces the semantic search space and gives agents a starting point.

Where alignment becomes insufficient

A candidate match is not the same thing as a trustworthy agreement.

There are at least four reasons for this.

1. Context can change the meaning

A term may look similar across two ontologies but behave differently in context.

For example:

  • rate may mean a monetary amount in one ontology
  • rate may mean a ratio or ranking in another

The same lexical signal can support multiple interpretations. Alignment alone does not always resolve that ambiguity.

2. Confidence scores are not decisions

A matcher may return:

price -> rate (0.71)

That score is useful, but the system still needs policy.

  • Is 0.71 enough to accept automatically?
  • Is domain risk low enough for a shortcut?
  • Should the agents request definitions first?
  • Should the system ask for evidence?

In STRATON-LLM, confidence is treated as an input to decision-making, not as the decision itself.

3. Real systems need challenge, not just suggestion

A good interoperability mechanism should allow agents to say things like:

  • “That mapping is plausible, but define your term.”
  • “Your concept only partially overlaps with mine.”
  • “Show the constraint that makes this correspondence valid.”
  • “I reject this equivalence but accept a narrower relation.”

Those are negotiation acts, not alignment outputs.

4. Weak agreements should be filtered out

Even if agents converge on a mapping, the dialogue can still be poor.

Examples:

  • one agent dominates the conversation
  • the other provides no meaningful counterargument
  • the agreement happens too fast
  • confidence fluctuates heavily
  • the same unsupported claim is repeated

STRATON-LLM handles this with a separate evaluation layer that scores the dialogue itself.

The implementation consequence

Once you accept that alignment is necessary but incomplete, the system architecture changes.

Instead of a one-shot alignment pipeline, STRATON-LLM uses:

| Layer | Purpose | | --- | --- | | Matching support | Produce candidate semantic links | | Trigger classifier | Decide whether negotiation is necessary | | Strategy selector | Choose the dialogue mode | | Proposal generator | Turn the strategy into structured acts | | Protocol engine | Control the turns and legal transitions | | Evaluation layer | Judge trustworthiness of the agreement |

This is how the project moves from candidate matching to operational semantic interoperability.

The practical lesson

If a system accepts alignment outputs too early, it creates a dangerous illusion of understanding.

The system appears to work because it always returns something. But in cross-domain environments, returning something is not enough. The mapping must be contextually justified and trustworthy.

That is why STRATON-LLM does not stop at alignment. It uses alignment as the opening move, not the end of the game.