Where the ideas are developed
Short, self-contained position pieces on why knowledge is a cross-cutting layer in agentic systems, why it tends to fail first, and what it means to engineer for it. The topics are meant to be read by a technical leader who is funding or building agentic work.
One idea, several angles
The flagship argument — that knowledge is a layer, not a module — is developed across a small set of topics. Each one below opens with the direct answer to why it matters, then goes into enough substance to be useful on its own. The topics are meant to be read by a technical leader who wants a clearer vocabulary for the knowledge layer.
Research topics
Why Knowledge for Agentic AI
The core argument: knowledge is a cross-cutting layer in agentic systems, not a single module.
- Direct answer
- Because agents act on what they know. A chatbot can be wrong in a sentence; an agent can be wrong in a decision, an API call, a record change, or a plan that keeps going because nothing told it to stop.
- Why it matters
- The knowledge layer is what says what is true, what is current, what the system is allowed to do, and what to do when it does not know. Neglect that layer and the model's fluency turns into confident action on bad information.
- Read
- Why Knowledge — full position piece
Knowledge in Enterprise Context
The hard part at organizational scale: collecting knowledge that lives in documents, tickets, chats, heads, and tools, and keeping it current enough that agents can trust it.
- Direct answer
- At enterprise scale, the knowledge problem is not ignorance — it is scatter, staleness, and ownership. The facts are in the building; they are just not assembled, current, or governed in a way an agent can safely act on.
- Why it matters
- Enterprise agents fail less from not knowing enough than from acting on the wrong slice of what is known: stale procedure, unowned data, mixed permissions, no deprecation. The work is assembly and governance, not invention.
- Read
- Knowledge in Enterprise Context
Knowledge in Personal Assistant Context
The hard part at human scale: a personal assistant that knows your emails, calendar, messages, and habits raises privacy, control, and legal questions.
- Direct answer
- A personal assistant is only useful if it knows you. The moment it does, the questions become: who else can see that knowledge, who controls it, how long it lasts, and what legal regime it falls under — DPDP in India, GDPR in Europe, and the rest.
- Why it matters
- Personal knowledge is the most valuable and the most sensitive knowledge an agent can hold. Memory across short, medium, and long term is useful; memory without control and compliance is a liability.
- Read
- Knowledge in Personal Assistant Context
Context Engineering
Context windows are a knowledge-delivery problem. This page treats context as something you design for, not something you hope the model will make sense of.
- Direct answer
- A context window is not a brain. It is a delivery mechanism for the right knowledge at the right moment. Treating it as one big bucket of text is the most common way to make an agent both more expensive and less dependable.
- Why it matters
- How you pack context determines what the agent can do, what it cannot do, how much it costs, and how often it acts on the wrong thing. Context engineering is the applied form of the knowledge-layer argument.
- Read
- Context Engineering — full position piece
Token Savings & Grounding
Where token cost actually comes from in agentic systems, and how grounding claims in retrievable knowledge reduces both cost and the kind of confident error that hurts trust.
- Direct answer
- Token cost in agentic systems is not primarily a model-cost problem. It is a knowledge problem: moving the wrong knowledge, re-reading it, and grounding nothing in anything you can point back to.
- Why it matters
- The same discipline that makes an agent cheaper — sending less of the right knowledge — also makes it more dependable. Grounding is the flip side of token savings: both are ways of taking knowledge seriously.
- Read
- Token Savings & Grounding — full position piece
How to use this research
If you are picking up this site for the first time, start with Why Knowledge. It is the flagship argument and the vocabulary the rest of the site builds on.
If you are working in a specific context, go straight to the piece that matches it:
- Why Knowledge — the core argument and the vocabulary.
- Enterprise — if your agents act on organizational knowledge.
- Personal Assistant — if your agents act on personal knowledge.
- Context Engineering — if you are designing how context is packed and delivered.
- Token Savings & Grounding — if cost and trust are the pressure you feel right now.
Each piece is written to stand alone. You do not have to read them in order, though the argument builds.
In this section
Agent systems
How agents are put together, and how they talk to each other and to tools.
- AI agents — what an agent is, and what changes when it acts rather than answers.
- A2A communication — agent-to-agent handoff, and what has to be agreed before it works.
- MCP — the tool and context interface that most agent stacks now speak.
- Agentic AI — planning, tools and side effects: why a wrong action costs more than a wrong answer.
Knowledge at work
Where knowledge has to survive contact with an organization and a person.
- Enterprise knowledge — the organizational version of the problem.
- Memory — the part of the knowledge layer the system writes itself, and why write policies decide whether it helps.
- Personal assistants — knowledge held for one person rather than an organization.
- Context engineering — choosing what the model sees on this call.
- AI context platforms — the analyst category for constructing, governing, and delivering knowledge to agents.
Frontier
Work that is still research, and is marked as such.
- Semantic communication — whether a shared model of meaning reduces what has to be transmitted.
- World models — systems that learn the structure of space and time, and the several different things the term covers.
Page updated .