Claim 2 · Active

Architecture becomes ontological, not just structural

Claim

The primary architectural work shifts from pipelines and components to shared ontologies: vocabularies, invariants, constraints, and mental models.

Rationale

Without shared meaning, faster execution increases divergence and coordination cost.

Supporting Signals

  • Need for registries, schemas, and canonical definitions
  • Emergence of planning planes (intent, decision, execution, verification)

Challenges

  • Risk of over-formalization
  • Cultural resistance to explicit modeling

Evidence

  • Overview • Ontology — Palantir Foundry documentation. Web page — requires manual capture (403 on automated fetch).
  • Ontologies • Overview — Technical documentation on building ontologies. Web page — requires manual capture (403 on automated fetch).
  • Ontology SDK (OSDK) • Overview — Developer SDK for ontology-driven applications. Web page — requires manual capture (403 on automated fetch).
  • Spec-Driven Development: From Code to Contract in the Age of AI — Academic paper formalizing spec-as-source workflows. Connects to Design by Contract. Positions vocabularies, invariants, and constraints as the 'real' system with code as a compiled by-product. First Tier A source for claim 2. LLM coding agents benefit immensely from SDD because structured specs reduce ambiguity.
  • Tools - Model Context Protocol Specification — Primary source: the actual MCP specification for tool definitions. Tools enable models to interact with external systems via a shared schema that agents discover and invoke. Effectively a machine-readable ontology of capabilities.
  • Spec-driven development | Technology Radar — Thoughtworks Technology Radar entry for SDD as an emerging technique. Different from the blog post — the Radar carries broader adoption signal. Teams maintain specifications, constitutions, and schemas as primary artifacts.
  • Model context protocol (MCP) - OpenAI Agents SDK — OpenAI's Agents SDK implementing MCP. Demonstrates cross-vendor convergence on shared ontological schemas for agent-tool interaction. MCP standardizes how applications expose tools and context to language models.
  • Unifying Large Language Models and Knowledge Graphs: A Roadmap — TKDE reference. LLMs are black-box and fall short on factual knowledge; knowledge graphs explicitly store structured facts. Roadmap for integration. Supports claim that explicit semantics and structured knowledge become more valuable alongside LLMs.
  • OWL 2 Web Ontology Language Document Overview (Second Edition) — W3C standard. Defines ontology concepts with formally defined meaning — classes, properties, individuals in RDF ecosystem. The cleanest reference for what 'ontological architecture' actually means vs. hand-wavy vendor language. Foundational.
  • A survey on augmenting knowledge graphs (KGs) with large language models (LLMs): models, evaluation metrics, benchmarks, and challenges — Peer-reviewed survey. Frames paradigms: KG-augmented LLMs, LLM-augmented KGs, synergized frameworks. Adds breadth and taxonomy for hybrid semantic architecture. Supports need for explicit ontological structure alongside LLMs.
  • Decentralized Multi-Agent Systems with Shared Context (DeLM) — DeLM achieves up to 10.5 points over the strongest baseline on SWE-bench Verified at roughly 50% lower cost by sharing verified context. Partial support for the ontology claim: a shared-verified-context substrate maps to, but does not fully instantiate, the broader ontology framing.
  • Ontology-Constrained Neural Reasoning — Ontology injection produces significant reasoning gains (p<.001 across 1,800 runs and 3 LLMs). Downgraded A to B: unreviewed preprint plus vendor self-evaluation of the authors' own FAOS platform (material conflict of interest).
  • Single Agents Match Multi-Agent Systems Under Equal Token Budgets — Single agents match or beat multi-agent systems when token budgets are held equal. Re-scopable challenge: cuts at multi-agent orchestration economics, not the ontology mechanism; the appeal to a processed version of context arguably argues for shared invariants.
  • A Strong Single-Agent Baseline for Multi-Agent Workflows — A strong single-agent baseline matches multi-agent workflows. Re-scopable challenge with the same caveat as Tran & Kiela: it cuts at multi-agent orchestration economics, not the shared-ontology mechanism.
  • Ontologies, Context Graphs, and Semantic Layers — Surveys production ontology instances (SNOMED CT, Gene Ontology, Siemens) as shared semantic substrates. Contextual support that ontologies operate as live architectural primitives in real systems.
  • Four Arguments Ontologists Never Finished — Catalogs central-but-unsolved tensions in ontology design. Contextual: surfaces the open questions in ontological architecture rather than resolving them.

Evaluation Criteria

  • Frequency of semantic vs technical disagreements
  • Stability of interfaces and concepts over time