The Agent I Didn't Mean to Build
A week ago I set out to build a tool. A small thing — a chatbot on my website that could answer questions about my career and work. Technically interesting, certainly. The capability to do this well, with retrieval over a personal corpus and a model that can actually hold a voice, did not exist in production form even two years ago. I wanted to see what it felt like to ship one.
What I ended up with is not a tool. It is an agent. I called it that on purpose, and the more I worked on it the more the word mattered.
From bullet points to narrative
The first version was built the obvious way. I gave it my resume, my LinkedIn, a stack of project descriptions. It answered questions. It was fine. It was also, quietly, embarrassing — because what it produced sounded exactly like every other AI summary of a senior technologist's career. Verbs about leading and architecting. Numbers in the right places. Nothing that was actually me.
So I started feeding it the rest. The arcs that connect the bullet points. The reasons I moved between roles. The texture of what it was like to lead an analytics organization, and what was different about advising a decade later. The story underneath the resume.
That is where the project stopped being a chatbot.
Why "agent" is the right word
"A chatbot represents an interface. An agent represents someone."
When I sat with the difference, I realized I had been quietly building a representation of my professional self that was richer than any artifact I had ever produced about my work. Richer than my resume — obviously. Richer than my website. Richer, in some respects, than how I would describe myself in conversation, because the agent contains the second-order context I rarely have time to lay out: why this engagement mattered, what I learned from it, how it connects to what I am doing now.
It is a digital twin in the only sense of that phrase I have ever found useful. Not a literal copy of me. A persistent representation of professional persona, voice, and judgment, available to interact with the world on my behalf when I cannot.
The shape of the world this points to
This is the part that shifted my thinking.
For most of the web's history, we have organized Information around documents and pages — static artifacts that other humans (or search engines) retrieve and read. The implicit model is library-shaped: indexed, hierarchical, retrieval-based. You go look something up.
An agent is shaped differently. It does not wait to be retrieved. It interacts. It answers, asks, qualifies, refers. It represents its principal in a transaction. When I imagine a web where every consequential professional has an agent like this — and every organization, and every program — the resulting environment looks far more like the physical world than the documents-and-search world we have spent thirty years building.
"It is a world of interactions, not lookups. Agency, not pages."
That is a more honest model of how knowledge work actually moves. Most decisions in my field are not made by reading a document. They are made through a conversation with someone who has read many documents and synthesized them into judgment. Agents move us closer to that shape — not as replacements for the humans, but as scaffolding around them.
The new problems this creates
I have been in this work long enough to be guarded about it. An agentic web creates real problems we have not solved.
Interactions compress search and qualification into the same moment, which is efficient, but they also consume foundational resources — bandwidth, compute, attention — at rates the document web never required. Multiply a single agent-to-agent exchange by every meaningful business interaction in a year and the infrastructure question becomes serious in a hurry.
It also breaks our current indexing assumptions. You cannot crawl an agent the way you crawl a page. You cannot rank one against another using PageRank-style heuristics. The federation problem — how agents discover, trust, and qualify one another — is genuinely open. New indexes and new hierarchies will have to be built. We are not at the end of search; we are at the beginning of something else.
And there is the harder question, which I do not yet have a clean answer to: what is the right relationship between a principal and their agent? Where does the agent's authority end? How do I update it without losing what makes it mine? Those are governance questions, and they will land on enterprises and federal agencies — and on the Systems that hold them together — long before they land on individuals.
A week is a long time
A week ago I set out to make a tool to curate my interactions with the world. In a few days I made something that shifted how I think about the next decade of this work.
"The agent is still small. The implications are not."
If you want to talk to it, it is on this site. Ask it something hard.