Frank Coyle’s talk is essentially a critique of today’s excitement around agentic AI. His central thesis is surprisingly simple:
LLMs can generate plausible actions, but without an ontology they do not really know what world they are acting in.
He argues that most current AI agents are built on probabilities (“the next likely action”), whereas real autonomous systems also require explicit knowledge about what things exist, how they relate, and what actions are allowed. Ontologies provide that missing layer.
He develops this argument in roughly five steps.
1. LLMs are probabilistic, not semantic
Modern LLMs are extraordinary statistical engines.
They know that
- invoices are often paid,
- customers place orders,
- refunds follow purchases,
because those patterns appeared during training.
But they do not inherently know what an invoice is, what distinguishes it from a receipt, or that one invoice cannot simultaneously be both “paid” and “unpaid.”
That distinction must be represented somewhere outside the language model.
2. Ontology is not a database schema
One of Coyle’s main points is that people confuse ontology with database design.
A SQL schema says
Customer
Order
Invoice
Payment
An ontology instead says
Customer
places
Order
Order
generates
Invoice
Invoice
settled_by
Payment
plus rules such as
- every payment belongs to exactly one invoice
- a cancelled order cannot be shipped
- an employee cannot approve his own reimbursement
These are meaning, not storage.
As several recent enterprise AI articles also emphasize, ontology describes how the business itself works rather than how tables are stored.
3. Agents need constraints
Today’s agent stacks often look like
LLM
↓
Tool call
↓
API
Coyle suggests inserting another layer:
LLM
Ontology
Validator
Tool
Now every proposed action can be checked.
For example
LLM says
Refund order #182.
Ontology checks
- Does order exist?
- Was it already refunded?
- Is refund allowed?
- Is requester authorized?
Only then does the API execute.
This makes the ontology a kind of guardrail rather than merely documentation.
4. Knowledge graphs are not enough
Another interesting distinction he makes:
Knowledge Graph
→ stores relationships
Ontology
→ defines what relationships are legal
Think of a road map.
A knowledge graph says
Taipei ---- Taichung
An ontology additionally says
- cities are places
- highways connect cities
- rivers cannot legally connect to customers
- one passport belongs to one person
In other words
Knowledge Graph = facts
Ontology = rules governing facts
5. Formal logic still matters
One thing I appreciated about his talk is that it pushes back against the current belief that “LLMs replace symbolic AI.”
Instead, Coyle argues something closer to
Probability
+
Logic
=
Reliable Agents
Language models remain excellent at
- interpretation
- planning
- dialogue
- ambiguity
Formal ontologies remain excellent at
- consistency
- constraints
- explainability
- auditing
The future is likely a hybrid architecture rather than one replacing the other.
Why this matters
His broader message is that today’s agents often sound intelligent because they speak fluently.
But speaking fluently is different from understanding the world they inhabit.
An ontology gives the agent a machine-readable model of reality:
- what entities exist,
- how they relate,
- what actions are permitted,
- what counts as an error.
Without that, an agent is essentially improvising from statistical associations. With it, the same agent can reason within explicit semantic boundaries, producing decisions that are more consistent, explainable, and trustworthy.
A connection to your own work
Reading your recent reflections on analogical self vs. algorithmic self, I found an unexpected resonance.
An ontology is, in a sense, an attempt to freeze a world into explicit categories and relations. It says:
“These are the kinds of beings that exist, and these are the legitimate relations among them.”
That is extraordinarily powerful for enterprise AI.
But from a Heideggerian or Deleuzian perspective—the thinkers you’ve often been exploring—it also reveals a limitation.
Ontology is excellent at describing a stable world.
It is much less comfortable with a world characterized by becoming, event (Ereignis), lines of flight, or novel forms of existence.
In other words, Coyle asks:
How can an agent know what already exists?
Whereas many of your recent writings ask a different question:
How can a person remain open to what does not yet exist?
Those are not contradictory questions. They point toward two complementary architectures:
- Agentic AI needs ontology to avoid chaos.
- Creative human existence may require the capacity to step beyond any fixed ontology into the emergence of new meanings.
That, to me, is where the discussion becomes philosophically interesting. It is no longer only about better AI systems, but about the difference between reasoning within a world and bringing forth a new world.
