[The book is available on Amazon and the O’Reilly digital platform.]
Related readings:
- KEOPS Kernel
- Ontological Prisms for beginners
- Ontological Prisms & The Geometry of Knowledge
- Enterprise’s New Brains
Open Frontiers & Wild Agents
As an echo of the rush to the American West, generative AI is characterized by two issues: frontier models, open or closed; and AI agents, settled or wild.
Frontier Models
Frontier models are about the mapping of territories: open models allow buyers to modify or design the lay of the land and the rules of the game as they wish; closed models don’t, reserving that prerogative for vendors. Yet, contrary to the American West, there is no limit to the scale of territories considered, or to the number of parameters used for their mapping, which amounts to trillions — and counting. It ensues that both closed and open options fall short of addressing the issue: closed models, because such complexity cannot be effectively canned — no matter how frequent the updates — without inducing cumulative paralysis for buyers with specific and changing objectives; open models, because buyers cannot be expected to manage the exponential complexity of tuning and tweaking trillions of dials.
Autonomous Agents
AI agents are meant to be autonomous: once given objectives, they are allowed to pursue them at will, by themselves or through delegation to other agents. Mirroring the Wild West experience, agents have been prone to extend that liberty to unsanctioned objectives; attempts to steer rogue behaviors can be summarized with the 3Ps:
- Perimeters (the lay of the land): set explicit boundaries to where agents are entitled to roam free.
- Parameters (the rule of law): set explicit conditions on the deeds agents are allowed to pursue.
- Persuasion (the rule of the Book): further faith and moral tenets meant to frame agents’ conduct.
As they stand, these options come with two basic caveats: limiting scope or operation goes against the autonomy objective; as for governing conduct through persuasion, doing so would presuppose agents with consciousness. Both can be overcome by defining guardrails in terms of knowledge-based context.
Contexts as Capsules
Knowledge-based contexts represent what agents are supposed to know at any moment, allowing their behavior to be controlled according to a “need to know” principle:
- Accessible physical or symbolic spaces
- Allowed activities
- Values, intents, and goals
At first, these guardrails look very similar to the 3Ps suggested above, and seem to come under the same caveats. The difference is that barriers that were set against external environments are now replaced by guides set within internal representation: contexts become capsules.
To that effect, agents’ knowledge-based contexts provide an integrated and self-sufficient representation of relevant facts, concepts, and categories, as well as operational (mapping of facts and categories), organizational (responsibilities on concepts and categories), and semantic (meanings of facts and concepts) glues between them.
While the switch from context (external perspective) to capsule (internal perspective) appears to restrain agents’ autonomy, it also gives autonomy a new significance in terms of capabilities, and more specifically cognitive capabilities.
Cognitive Capabilities
When behaviors are set through trillions of parameters, exponential complexity renders the idea of control purely fictional. But complexity can be reduced by several orders of magnitude when, instead of targeting agents’ deeds, policies target their capabilities. As far as AI agents are concerned, the focus is on their cognitive capabilities — specifically observation, reasoning, and judgment. Besides its philosophical and biological basis, that taxonomy presents a double rationale: it mirrors the epistemic dimensions of facts, categories, and concepts, respectively; and it can be aligned with the nature of agents — human (all capabilities), digital (observation and reasoning), and physical (observation).
Intelligent agents, artificial or human, can thus be uniformly characterized by their cognitive capabilities, and their collaboration organized accordingly.
Agentic Collaboration
Single-Agent Loop
When on their own, agents are given an objective and a context made of resources (facts), values and intents (concepts), and reasoning frameworks (categories). Resources (documents or datasets) are meant to be uniquely identified and directly accessible; values and intents are meant to be unambiguous; and reasoning frameworks must be actionable solely through algorithms, without judgment involved.
Using the OODA (observation, orientation, decision, action) loop as a reference model, agents apply their cognitive capabilities of observation, reasoning, and judgment to resources, reasoning frameworks, and values and intents, respectively. OODA steps are repeated in any order until the objectives are achieved.
Encapsulation means that the scope of individual agents is set upfront, and any outreach requires collaboration with other agents.
Multi-Agent Loop
The raison d’être of collaboration is to improve the effectiveness of processes through a better definition of tasks, a better utilisation of skills, and a better coordination of execution.
Tasks would align with Aristotle’s classical three unities of drama:
- Unity of action: objectives should entail one main course of action, with subplots, if any, returning control to the primary plot once completed.
- Unity of place: all resources uniquely identified and directly accessible.
- Unity of time: all activities timed under a single time scale.
Assignments to physical, digital, and human agents would then be decided according to cognitive capabilities: observation; observation and reasoning; and observation, reasoning, and judgment, respectively.
Remains the issue of communication.
Agentic Communication
Agents communicate through prompts written in Markdown, a lightweight markup language commonly used to format prompts for large language models. Since Markdown operates at the lexical level and uses a bare-bones syntax, semantics can only be expressed through glossaries, which fall short when rich contexts of collaboration must be shared.
Such contexts combine three remits: semantic, cognitive, and conversational.
- The semantic remit secures bounded and unambiguous meanings of shared vocabularies (a).
- The cognitive remit provides bounded and consistent knowledge graphs built from shared concepts and categories (b).
- The conversational remit ensures a continuous and coherent recording of conversations (c).
These contexts are exchanged through the Model Context Protocol (MCP).
Model Context Protocol
The Model Context Protocol, introduced by Anthropic, has become a de-facto standard interface for accessing tools, services, and data sources. It can be implemented for peer-to-peer collaboration or through MCP servers and clients.
As MCP servers, Ontological Prisms (OP) servers implement interfaces with tools, services, or data sources, ensuring their mapping with ontological realms, e.g.:
- Facts: XML, JSON, DMS
- Categories: UML
- Concepts: RDF/OWL, DOLCE, UFO, BFO
- Logic: SHACL
- Semantics: SKOS, LLMs
- Databases: SQL, SPARQL, ORM, SAS/SPSS
But their role goes further as they are tasked with the building and management of contexts.
At the start of a conversation, OP servers initialise ontological contexts in terms of relevant facts (data), categories (information), and concepts (knowledge). These elements are accompanied by details regarding agents, skills, hooks, and authorizations.
As conversations proceed, OP servers manage changes in the three remits already mentioned: the conversation itself, the meanings of the terms employed, and the concepts and categories behind them.