GraphNews

Why agentic AI fails on meaning instead of plumbing and what your organisation must own to close the gap.

Why the Feedback Loop that Built your Product is The Same One that Will Wreck It

Vector databases find isolated text. Knowledge graphs map reality. Here is the blueprint for the next era of AI architectures.

What the shift in AI agent architecture is really about

Two variables, one change, and what a backlink is actually for

Knowledge → Semantics → Action

Urgency of bounded context and the architectural quantum of data that enables it.

If you’re building a company brain, GBrain is required reading. Its 6 layers explain the machinery and reveal where the company problem begins. GBrain is Garry Tan’s open-source brain layer for his AI agents. It ingests meetings, emails, voice calls and notes, connects them, and gives his agents memory that survives beyond one chat. Basically, one layer feeds the next. 1/ The knowledge graph turns flat notes into entities and relationships. 2/ MCP connects the agent to the external tools where information lives. 3/ Skills tell the agent what to do with those tools. 4/ Hybrid search combines semantic meaning with exact keyword matching to find the right material. 5/ Agentic RAG decides how to retrieve that material, checks it, and uses it to build an answer. 6/ A company brain adds the maintenance layer that keeps the whole system current, trusted and usable. My take is simple. GBrain is a very good blueprint for a personal brain. The company brain problem gets quite a bit harder. Multiple people bring different permissions, conflicting sources, stale knowledge, and updates that need clear ownership and approval. That is the gap between personal memory and shared company truth. Still, GBrain is one of the best repos to play with if you want to understand how the plumbing works. https://lnkd.in/erKGf4tY

Product information management was built to describe products. AI readiness requires a system that defines them.

Discover how we modeled our MongoDB infrastructure as a traversable graph and then use pathfinding algorithms to dynamically compute and execute recovery plans. This automated approach reduced pager volume by 30% (~200 pages/year), eliminated 12 days of unhealthy shard states annually, and supports new shard layouts with zero manual effort.

Following up on my previous post outlining the baseline requirements for a foundational (top-level) ontology, here is Part 2. As a reminder, we are avoiding the vague "requirement" concept by using the SysFEAT triad: the Job-to-be-done, the Capabilities required, and the overarching Principles. Here are the final four pillars required for an ontology to succeed in the future of Architecture and AI: 5. Granularity & Levels of Reality 🎯 Job: Model complex domains spanning multiple scales and abstraction layers. 🛠️ Caps: Multi-dimensional granularity (conceptual, systemic, spatial, and organizational levels). See Matthew West document on level of reality for illustrating: https://lnkd.in/eam59uWS. ⚖️ Princ: An entity's level dictates its applicable properties. Shifts across levels must be logically mapped, never conflated. 6. Language & Culture Neutrality 🎯 Job: Ensure global usability across cognitive models. 🛠️ Caps: Strict decoupling of the conceptual model from lexical or cultural representations. ⚖️ Princ: Model fundamental reality/structures, not regional natural language semantics. 7. Property Binding & Value Spaces 🎯 Job: Quantify and qualify entities consistently. 🛠️ Caps: Formal patterns mapping entities to properties and value spaces. ⚖️ Princ: Strictly define property binding and measurement resolution to prevent ambiguity in data integration. 8. Metaphysical Transparency 🎯 Job: Prevent modeling errors and facilitate peer evaluation. 🛠️ Caps: Comprehensive documentation, competency questions, and explicit philosophical choices. For illustration: From BORO (Chris Partridge): A couple of meta-ontological choices for ontological architectures - https://lnkd.in/eD2cMkXT From UFO (Giancarlo Guizzardi): Towards Ontological Foundations for Conceptual Modeling - https://lnkd.in/epBzHu5E ⚖️ Princ: Total transparency regarding formal and metaphysical stances is mandatory. In the next post, we will explore item 1 about syntax and modularity. (If you missed Part 1 covering Syntax, Identity, Relations, and Computability, check my previous post! Let me know what you think of this baseline in the comments).

Microsoft CEO Satya Nadella warns that a handful of AI models could hollow out entire industries, as Microsoft faces rising AI costs, enterprise pressure and questions over who will control the value of artificial intelligence.

We are entering an era where AI agents have officially moved from acting as passive assistants to autonomously owning decisions. From incident response to credit approvals, agents now make recommendations and coordinate work across complex enterprise systems. However, this profound shift exposes a critical new bottleneck: context.

The ontology schema below illustrates a Minimally Interoperable Profile (MIP) for the food and health domains. It reuses established ontologies such as FoodOn, HL7 FHIR, Schema.org, PROV-O, DPV, and DUO to define the minimum shared semantics needed for interoperability. This is the real power of ontologies in interdisciplinary research. Different communities can continue using their own data models while agreeing on a common semantic layer that allows their data and metadata to be represented as interoperable knowledge graphs. The result is data that is easier to integrate, discover, query, and reuse—without requiring complete harmonization of every dataset. As research increasingly depends on combining data across food, health, agriculture, biodiversity, environment, and many other domains, lightweight semantic profiles such as MIPs can provide an accessible starting point for FAIR data, interoperable data spaces, and AI-ready knowledge. Sometimes, connecting disciplines starts with agreeing on just a handful of shared concepts. #KnowledgeGraphs #Ontologies #SemanticWeb #FAIRData #Interoperability #FoodData #HealthData #KnowledgeEngineering #LinkedData #DataSpaces #AIReadyData #FoodOn #FHIR

Why our team chose the RDF stack (RDF, OWL and SHACL) instead of building our own semantic framework.

A neuro-symbolic AI agent that remembers in a knowledge graph and reasons with an OWL engine — consistent, multi-hop, and auditable by construction.

Every enterprise racing to deploy AI agents is hitting the same wall: the models are brilliant, but they don't know what your business means by "customer," "exposure," or "active contract." They hallucinate not because they're broken, but because nothing in your architecture tells them what's true.

🧠 Prompt Engineering → 🗂️ Context Engineering → 🔁 Loop Engineering → 🕸️ Graph Engineering Four eras. 24 months. If your team is still optimizing for the first one, you're not behind ... you're solving last year's problem with this year's budget. On July 18, developer Peter Steinberger, the guy behind OpenClaw, asked a six-word question that reframed the whole debate: are we still talking loops, or did we shift to graphs? Half a million views in hours. Not because it was profound. Because a lot of engineering leaders were feeling it and hadn't said it out loud yet. I know that feeling. You finally get your arms around one operating model for AI-assisted development, and the ground moves again. If you're the one explaining this to your board next quarter, that whiplash is real. Here's the pattern underneath the buzzwords and why it matters for how you architect this work: 🔁 A loop is one agent, cycling through a task, deciding for itself what's next. Forgiving to set up. But a loop is a regulator, not a reference-setter it can hit a target you've already defined (tests passing, a clean build), but it can't decide what "correct" means. Point it at an ambiguous spec and it doesn't fail cleanly. It burns tokens guessing, in circles, until someone notices the invoice. 🕸️ A graph forces that decision up front. You map explicitly which agent owns which domain, how data hands off, what happens when a step fails. Harder to set up. Far more predictable at scale. What I'm seeing in production-minded teams: a split into two layers. 🏛️ A permanent org graph: named agents that own security, data, API contracts, frontend standards, and defend those boundaries continuously. ⚡ An ephemeral work graph: task nodes spun up for one feature or bug, that split, merge, and dissolve once validated, handing the result back to the permanent owners. If you're deciding how to structure this in your own org, three things I'd hold the line on: ✅ Keep shared state flat and strictly typed a bad output should route to a fix-it step, not corrupt everything downstream. ✅ Separate the model that generates code from the logic that decides what happens next. You do not need the same model you dev with to be your agent. ✅ Bound everything, no token maxing, iteration caps, token ceilings, a runtime limit, and a rule that pages a human the moment the state stops changing. That last one isn't optional. 📊 The FinOps Foundation's latest report found 98% of enterprises are now actively managing AI-related costs, up from 31% two years ago. That jump is what unbounded agent loops look like on a P&L. The technology under this shift is genuinely interesting. But the leadership question hasn't changed: where are the humans still setting the reference and where have you quietly let a system start guessing? #AI #EnterpriseArchitecture #DigitalTransformation #AgenticAI #CIO

Kurt Cagle & Chloe Shannon | The Ontologist

Constants, fluents, and the quiet failure of reification

Over the past several months, I have been building AICHAX — the trust layer for AI, where Answers are computed, not generated. LLMs hallucinate by design — it's built into the architecture. AICHAX takes the opposite path: a Graph Intelligence Engine where every answer is the deterministic result of traversing a real knowledge graph. Sourced. Auditable. Instant. And when the graph doesn't know — it says so, honestly. In under five minutes, watch it: 🔍 Verify a famous Einstein quote back to the original 1926 letter — source on screen 🧮 Compute math deterministically — steps shown, answered in milliseconds 🧬 Walk real relationship chains from Alan Turing to Geoffrey Hinton — live ⏳ Reason counterfactually over documented influence — evidence, not invented essays 🕸️ Flip any answer into a living, explorable graph — the evidence IS the interface 👪 Build sourced family trees, browse the painting canon, spiral ten thousand years of art into a galaxy 🧠 Ask your own mind — a Personal Wisdom graph that cites your own saved knowledge back to you Hallucinations end where structure begins. https://lnkd.in/gw9gAcjZ

I wrote an article, where I offered this definition of context: “Context is the situation or conditions that constrain both what information is relevant and how it should be interpreted to support effective decisions and actions.” In response, many people told me their knowledge graph provides their

Sign in or join now to see posts like this one and more.

𝙏𝙝𝙤𝙪𝙜𝙝𝙩 𝙛𝙤𝙧 𝙩𝙝𝙚 𝙙𝙖𝙮: There’s a new layer taking shape in AI and agentic AI solutions: the 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗟𝗮𝘆𝗲𝗿 To me, 𝙩𝙝𝙞𝙨 𝙞𝙨 𝙩𝙝𝙚 𝙣𝙚𝙬 𝙄𝙣𝙛𝙤𝙧𝙢𝙖𝙩𝙞𝙤𝙣 𝘼𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚 𝙛𝙤𝙧 𝙩𝙚𝙘𝙝𝙣𝙞𝙘𝙖𝙡 𝙙𝙤𝙘𝙪𝙢𝙚𝙣𝙩𝙖𝙩𝙞𝙤𝙣! It may seem relatively new, though some of us have spent the past several years warning that it was coming - in the form of taxonomies, ontologies, and knowledge graphs. The technical documentation community is in danger of giving away its future. This layer should belong to Information Developers and Information Architects, in partnership with engineering. It is not a side concern or an adjacent discipline. It is the next logical extension of information architecture - and one of the richest opportunities for new influence, new roles, and next-generation careers. Yet much of the profession remains focused almost exclusively on source content and source content structure. That work matters, but it is only half of the emerging information architecture. Organizing content without owning the knowledge layer, which is the extension of content structure, is what makes it discoverable, teachable, and usable is like designing a building and then handing over the floor plan to someone else to decide how people will move through it. If technical communicators do not take control of the knowledge management layer built on their content, developers will. And when that happens, the profession will not simply lose ownership of a system. It will surrender the very jobs, authority, and strategic relevance it has spent years claiming it deserves. | 49 comments on LinkedIn

The Enterprise Knowledge Layer (EKL) is a shared, governed substrate where an enterprise's knowledge lives. Learn more about it in our blog.

A library catalog is a deceptively hard graph problem. It looks simple: books, authors, subjects. But the moment you try to model derivations you're in deep waters. For instance, this translation was made from that critical edition, which itself corrects an earlier manuscript. That's before you touch provenance, uncertainty, or the fact that "Shakespeare" as an entity means something different in 1623 versus today. The Hamlet that existed as a quarto in 1603 is a different Instance than the First Folio version of 1623, which differs from the modern critical edition... The RDF bet: semantics first. BIBFRAME, the Library of Congress's replacement for MARC, is built on RDF. Its core move is to separate Work (the abstract creation) from Instance (a particular published form) from Item (the physical copy on the shelf). These aren't just categories, they're nodes in a graph, connected by typed predicates. RDF gives you something LPG cannot easily match: global identity. The cost is expressiveness at the edge level. RDF's triple model makes it verbose to attach metadata to relationships. When did this attribution change? Who contested it? RDF-star is addressing this, but it remains an afterthought on a model designed for facts, not so much claims. The LPG bet: richness at the edges. A labeled property graph lets you put properties directly on relationships. "Influenced by" can carry a weight, a date, a confidence score, a provenance note. Reification gymnastics is out 👌. For a museum modeling the relationship between an artist and a movement, or a library tracking the editorial history of an attribution, this is genuinely more natural. The tradeoff is the loss of global semantic grounding. LPG has no native concept of shared identity across systems. You can build it but it's not given to you by the model. Every institution ends up with its own node for Shakespeare, and reconciliation becomes a pipeline problem rather than an architectural guarantee. Cultural heritage institutions that care about interoperability (e.g. Europeana, the DPLA, national libraries) seem to have converged towards RDF and linked open data. The semantic web vision of a global cultural graph only works if nodes mean the same thing across systems. Visualization and dashboarding (BI) atop RDF is more complex than LPG. How this happens in practice I have not researched. Would enjoy input from the experts here 🤔 - BIFRAME implementation: https://lnkd.in/ebssxenu - Library workflows: https://lnkd.in/emy7hwJ6 - Singapore Library Blog: https://lnkd.in/ebbWvmQ2 #LibraryTech #KnowledgeGraphs #GLAM #RDF

As promised, here is a breakdown of what a foundational (or top-level) ontology needs to succeed in the future of Architecture and AI. To be clear, this isn't presented as the final, authoritative list, but rather a robust baseline designed to cover the major topics at their best. Because this is a massive topic, I am splitting this framework into two posts. Today, we look at the first four pillars. Because "requirement" is such an overloaded term in systems engineering, I’m structuring this using the SysFEAT methodology. Every point below is defined by a strict triad: the Job-to-be-done, the Capabilities required, and the overarching Principles that constrain the solution. Here is Part 1 of the 8-point master framework: 1. Foundational Syntax, Modularity & Conservative Extension 🎯 Job: Establish a robust, modular scaffolding that houses baseline vocabulary and allows non-destructive growth. 🛠️ Caps: A formal syntactic layer isolating structure from semantics, with absolute lexical scope and local encapsulation. ⚖️ Princ: Syntax precedes semantics. Additions must be strict conservative extensions with zero silent side-effects. 2. Spatiotemporal Existence & Identity 🎯 Job: Track entities across their entire physical and temporal lifecycle. 🛠️ Caps: Formal absolute/relative 4D indexing mechanisms backed by strict criteria of identity. ⚖️ Princ: Identity must be consistently verifiable (extensional or intensional) across an entity's entire footprint. 3. Semantically Exhaustive Relations & Reflexive Classification 🎯 Job: Provide comprehensive structural, existential, and taxonomic links. 🛠️ Caps: Axiomatized mereology, topology, directional holonymy/meronymy, and higher-order reflexive classification. ⚖️ Princ: Relations must capture fundamental asymmetries. Classification must support dynamic, cross-order aspectual roles without rigid stratification. 4. Logical Rigor, Computability & Proof 🎯 Job: Guarantee automated inference and machine-verifiable correctness. 🛠️ Caps: Built on formal, decidable logic (e.g., Higher-Order Logic) supporting strict computability and mathematical proofs. ⚖️ Princ: Formal rigor is absolute. Unverifiable constructs are strictly invalid. (Stay tuned for Part 2 newt week, where we will cover Granularity, Neutrality, Property Binding, and Metaphysical Transparency!)

I am pleased to share that a set of #Gartner vendor scorecards for popular data management platforms landed today. 𝑾𝒉𝒚 𝒅𝒊𝒅 𝒘𝒆 𝒖𝒏𝒅𝒆𝒓𝒕𝒂𝒌𝒆 𝒕𝒉𝒊𝒔 𝒆𝒗𝒂𝒍𝒖𝒂𝒕𝒊𝒐𝒏? Client demand. Almost a half of Gartner client interactions between January 2025 and April 2026, specifically within the “data management solutions” perennial priority, involved one or more mentions of #Microsoft, #Databricks, #Salesforce, #Snowflake, or #Palantir. Some of these conversations were direct comparisons of these data management platforms. D&A leaders regularly inquire about the effectiveness of these platforms for meeting use cases such as data foundation, semantic enrichment, domain enablement and agentic automation. 𝑾𝒉𝒂𝒕 𝒂𝒓𝒆 𝒕𝒉𝒆𝒔𝒆 𝒔𝒄𝒐𝒓𝒆𝒄𝒂𝒓𝒅𝒔 𝒇𝒐𝒓? These scorecards provide a holistic assessment of platform capabilities. D&A leaders can find answers to two key questions: 1. 𝐇𝐨𝐰 𝐦𝐚𝐭𝐮𝐫𝐞 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐚𝐥 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬 𝐨𝐟 𝐭𝐡𝐞 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦?  This knowledge is essential for them to compensate for any specific platform weaknesses with targeted point solutions, ensuring that their data tech stack is comprehensive. 2. 𝐇𝐨𝐰 𝐞𝐟𝐟𝐞𝐜𝐭𝐢𝐯𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦 𝐢𝐧 𝐟𝐮𝐥𝐟𝐢𝐥𝐥𝐢𝐧𝐠 𝐬𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐮𝐬𝐞 𝐜𝐚𝐬𝐞𝐬? This understanding is essential for executing their D&A strategy, that is, knowing what they can achieve with their technology investment. We have identified four use cases for the data management platforms: data foundation, semantic enrichment, domain enablement and agentic automation. My sincere appreciation to several colleagues who were instrumental in this research including Ehtisham Zaidi, Rita Sallam, Adam Ronthal, Masud Miraz, Aaron Rosenbaum, Nina Showell, Mark Beyer and Zain Khan. Its a must-read research for all Gartner clients https://lnkd.in/eYNGBtpn . #GartnerDA #datamanagementplatforms