Behind the work
How this actually works.
I’ve built a small team of AI specialists to work alongside me. It’s a standing group rather than a single chatbot I open when I need something, and each one is set up for a particular kind of work. The whole thing runs on my own infrastructure.
It’s organized the way a real office is. A chief-of-staff agent sits at the top, takes whatever I hand it, and parcels the work out. Behind it are specialists, each with its own brief and its own memory. One is built around financial analysis and Canadian tax, another follows the markets, another drafts in legal language. A small engineering crew builds and maintains the software the rest of them rely on, and a librarian keeps the knowledge base in order.
The benefit, honestly, is time. The morning briefing is written and waiting before I’m up. Research that used to take an afternoon comes back in minutes. Several things move at once, and because each specialist remembers what came before, I’m not re-explaining context every time I sit down. None of it replaces my judgment. It clears the runway so I can spend my hours on the parts that actually need me.
What follows is a plain-English look at how the whole thing is put together. What the technology can really do, how the pieces fit, and the parts that are still rough. A tour of the machinery for anyone who’s curious, not a pitch.
01 / The shift
From chatbot to agent.
Chatbot
A single conversation. You ask, it answers, and the whole exchange lives in one text box.
Describes the work.
Agent
It plans the work, runs it, watches what comes back, and revises. Along the way it reads files, calls APIs, opens browsers, and stages emails for review.
Produces the work.
Sounds like a small change on the page. It plays out as a very different shape of work.
02 / Architecture
The right mental model is a small operating team.
A coordinator at the top takes the request and pulls in whichever specialists make sense. The specialists run alongside each other and return their pieces. The principal stays on the loop for review and the calls that matter. Illustrative. The actual bench is broader.
Specialists are narrow on purpose, so they can stay deep and run alongside each other. Their work funnels back through the coordinator before anything reaches the principal. A single chatbot conversation has no way to do that. A small operating team is built for exactly that shape.
03 / Memory
Memory that stays inside the perimeter.
A year ago, persistent memory was the differentiator. Today every major AI vendor remembers between sessions in some form. The interesting question is no longer does it remember? It is who owns the place the memories live?
Local
Memory lives on hardware I control. A small server in my office runs the vector database and the plain-text file store behind it. None of it flows through a vendor’s cloud by default.
Secure
Anything sensitive stays inside a perimeter I can actually describe and point to. Nothing reaches a consumer AI product by accident.
Portable
The whole memory layer is independent of any one AI vendor. The same accumulated context wires up to whichever model is best tomorrow.
04 / Tools
Most real work is a chain of tool calls.
How a single task actually runs.
- 01Pull data
- 02Reconcile
- 03Build model
- 04Assemble memo
- 05Render PDF
- 06File
- 07Stage email
A chatbot can do any one of those steps when asked. An agentic system runs all seven in order, on its own, and surfaces only the final deliverable for review. The connective tissue that lets it reach those tools is a standard called MCP, the Model Context Protocol, which is quietly becoming the way models talk to the systems around them.
05 / Autonomy
Agents don’t have to wait to be asked.
Scheduled
Runs on a clock.
- Morning briefing at 6:15 daily
- Weekly market scan on Friday afternoons
- Quarterly portfolio refresh
Event-driven
Runs on a trigger.
- Email from specific contact → draft reply
- Stock hits price threshold → notification
- Calendar invite accepted → briefing pack
The bench gradually moves from things I can ask for over to things that are simply getting done.
06 / The honest part
Anyone who tells you AI is solved is selling something.
- 01
Hallucination is real
AI invents plausible-sounding wrong answers. The way to keep that out of anything serious is to ground every claim in a retrieved source and put a human on the read before it goes anywhere.
- 02
Long tasks drift
Quality drops as conversations get long. The fix is to break work into smaller pieces, run them in parallel, and bring the results back together. That is the whole reason the specialist pattern exists.
- 03
No judgment on ambiguity
The model produces something. It has no way to know whether that something is right for you. The human stays in the loop on every decision that matters.
- 04
Compute is not free
Routine work runs on the cheapest model that can do the job. The expensive models are held back for genuinely hard reasoning. The economics work when the routing is done carefully, and they fall apart when it isn’t.
- 05
Privacy is a design decision
Sensitive material does not flow through consumer AI products by default. Enterprise terms, retention controls, and local infrastructure have to be designed in at the start. They are not a setting you flip on later.
07 / Where this leaves us
What this technology actually changes.
What changes for the better is real. Coordinating large amounts of information gets faster. High-quality drafts under time pressure stop being a heroic effort. Continuity across long stretches of work is easier to hold, and several things can run in parallel without any of them getting dropped. None of that is small.
What doesn’t change is the human work. Judgment and relationships still sit with a person. Accountability still has a name on it. An AI bench can produce a board memo, and it cannot sit in the boardroom.
The interesting part is the combination. Fast like a small team, continuous because it genuinely remembers, and still answerable to a person who reads the output. There is no magic in it. It is an architecture, and the architecture is worth understanding on its own terms.
None of this is secret, and none of it is for sale on this page. If you’re building something similar, or just curious how the pieces fit, I’m always happy to compare notes.
Say hello