Conventions anyone can adopt
Open standards and conventions for knowledge representation, lifecycle, grounding, the agent interface, and the discoverability layer that lets agent engines find and use knowledge reliably. The goal is a small, usable set of standards — not a catalogue.
What is this standards work for?
Most agentic systems do not fail because there is no standard for knowledge. They fail because every team quietly builds its own, and then the knowledge is inconsistent across systems, stale in places, permissioned wrong, and impossible to ground a claim in. A small, usable set of open conventions is the difference between "everyone approximates" and "the knowledge layer is dependable enough to act on."
Principles
A few standards, well-chosen
A standard that no one adopts is not a standard. This work favors a small set of conventions that cover the actual failure modes: representation, lifecycle, grounding, the agent interface, and discoverability.
Free to read, not gated
Anyone can read and use these pages — teams, open-source projects, non-profits, personal assistants — with nothing gated. A vendor's proprietary platform gets the same treatment as a W3C Recommendation. The org side exists so this work is free to read: not-for-profit, no fee, no ads. It carries platform material alongside open standards, because the commitment here is free reading, not an open licence for everything.
Usable before complete
Standards are adopted when they reduce pain, not when they are philosophically complete. Each convention below is written to solve a real problem today, and to leave room to grow without collapsing under its own weight.
Standards & platforms
Knowledge representation
How knowledge is shaped so that agents can use it without re-deriving the structure every time.
- Problem
- Knowledge lives in documents, tickets, chats, spreadsheets, heads, and tools. Each has its own shape. Agents that act on that knowledge need a consistent way to find, read, and trust it.
- Convention
- A small set of representation primitives — entities, facts, procedures, policies, and the relationships between them — with enough structure to be usable and enough flexibility to fit real content.
- Status
- Working draft; developed in the research pieces and refined through adoption.
Knowledge lifecycle
How knowledge is created, kept current, and retired — because knowledge that is not governed decays into confident error.
- Problem
- Knowledge is created somewhere, updated somewhere else, and retired never. Agents that act on stale or retired knowledge are the most expensive kind of wrong.
- Convention
- A lifecycle model with clear ownership, update cadence, deprecation, and the idea that "current" is a property an agent can check, not an assumption it makes.
- Status
- Working draft; the lifecycle ideas are developed in the research pieces and shaped through adoption.
Grounding
How an agent ties what it says or does to knowledge it can point back to — so it is more dependable and more auditable.
- Problem
- An agent that says something plausible but cannot show its basis is hard to trust in production. Grounding is the discipline of tying claims back to retrievable knowledge.
- Convention
- A grounding model where claims carry their basis: what knowledge was used, when it was current, and what the agent was allowed to do with it. The same discipline also reduces token cost by sending less of the right knowledge.
- Status
- Working draft; see the token savings and grounding research piece.
The agent interface
A clean, documented surface that agent engines can rely on to find and use knowledge — as opposed to scraping whatever is easy.
- Problem
- Agent engines today often have to guess how to get knowledge out of a site or system. A clean interface — a documented surface, a convention for what is available and how to ask — makes knowledge actable instead of merely visible.
- Convention
- A small agent interface convention: what knowledge is exposed, in what shape, with what permissions, and how an agent engine asks for it. This is where things like
llms.txtand a documented agent surface fit. - Status
- Working draft; see the free resources page for the discoverability layer.
Discoverability for agent engines
How knowledge is made findable by agent engines — the AEO / Agent Engine Optimization layer.
- Problem
- If an agent engine cannot find your knowledge, it cannot act on it. Discoverability is the difference between knowledge that exists and knowledge that is usable at distance.
- Convention
- A discoverability layer — a documented surface, a file agents can read to understand what knowledge is available and how to ask for it. This is the org-side counterpart to the search and SEO work that already exists for human readers.
- Status
- Working draft; see the free resources page.
How to use these standards
These conventions are not a checklist to implement all at once. They are a map of the failure modes and a set of starting points. Most teams will adopt one or two first — usually representation and lifecycle, because those are where the pain shows up first.
If you are adopting from zero, start here:
- Free resources — the public-facing starting points: llms.txt, the agent interface, and the discoverability layer.
- Adoption — how to put knowledge-first practice into place, including maturity evaluation and lifecycle adoption.
- Articles — shorter, more current thinking on what is happening in the field and what it means for standards.
The standards are developed openly and refined through adoption. If a convention does not work in practice, that is the feedback loop — not a reason to stop.
In this section
Discovery
Telling a machine what a site contains before it starts guessing.
- llms.txt — a short markdown map of a site, at a known path.
Knowledge organization
Controlled vocabularies and their relations, so two teams stop naming the same thing twice.
Data model
The layers that give statements structure and meaning.
- RDF — the triple model underneath everything else on this list.
- RDFS — classes and properties, with inference rather than constraint.
- OWL — formal ontology, its profiles, and the cost of reasoning.
- Open Knowledge Format (OKF) — Google Cloud's markdown knowledge bundles, cross-linked into an agent-walkable graph.
Serialisation and storage
How the model is written down, and the open file and table formats it lands in.
- Turtle — the compact RDF syntax people actually write.
- JSON-LD — linked data in the format web developers already use.
- HDT — compressed RDF for publishing and archiving at size.
- Parquet — the columnar file format, openly licensed, and the substrate most platforms read.
- Delta Lake — transactions and time travel over Parquet, open source.
- Iceberg — the open table format competing platforms adopted.
Query
Asking a graph a question, and getting an answer you can check.
- SPARQL — graph patterns, federation, and update for RDF.
- GraphQL-LD — querying linked data with GraphQL, translated to SPARQL.
Provenance and persistence
Where a claim came from, and what the source said before it changed.
Platforms
The proprietary platforms enterprise knowledge actually sits in. Who controls the schema is the axis here, not whether the code is open.
- Databricks — a proprietary lakehouse and its catalogue.
- Snowflake — a proprietary warehouse, and where permissions are enforced.
- Palantir — the Foundry Ontology, the closest commercial analogue to an ontology layer.
- ServiceNow — CSDM and the CMDB: a vendor-imposed governance model.
- Salesforce — a huge installed base, and a schema you cannot port.
- Microsoft Graph — the interface most enterprise agents actually read.
- Neo4j — property graphs and Cypher, against RDF triples.
Page updated .