Class 2
Managing AI Workers — Part 1#
The architect's capability stack — connectors (MCP), skills, and the System of Record — plus what's new in Claude.
Turning a plain chat model into a capable AI worker comes down to three moves: give it access to tools with a connector, teach it the procedure with a skill, and point it at a System of Record for the truth.
From Sir Zia’s class — Zia Tutor AI, The AI Agent Factory (FDE track).
Why it matters
If you can’t separate connecting a tool from teaching a worker to use it, you can’t design an AI worker. This class is the architect’s mental model for giving a worker real-world capability.
The capability stack
Connector
A bridge that gives an AI worker access to an outside system — an app, a database, a website — so it can read data and take actions there. In Claude you add one from the connectors panel with the Add button and the connector’s link, even connectors that aren’t in the built-in list.
MCP — Model Context Protocol
The open standard connectors are built on — think of it as the “USB-C port” for AI tools. It exposes tools (actions the worker can call) and resources (data it can read) to the model.
Skill
A reusable, written procedure that teaches the worker how to do a task the way you want — including how to use a connected tool. You can write one by hand, or use the built-in
skill-creator skill to generate it.
System of Record (SoR)
The authoritative store of truth for a domain’s data — a learner record, a CRM’s customer data, an inventory. The worker reads and writes it through a connector.
The key insight: connecting ≠ knowing how
My note
If I connect a piece of software to Claude, can it instantly use it the way I do? No. The connector gives access; the worker still has to learn the procedure. A skill is the best way to teach that procedure — it captures “the way I use this software.”
Where the System of Record fits
I wasn’t sure how the layers stacked. The corrected model:
- Software / data you want the worker to use — this may itself be the SoR (the authoritative data).
- Connector (MCP) — the wiring that gives the worker access to that software or data.
- Skill — the know-how: the procedure for using it well.
Correction
SoR = System of Record (often called the source of truth). It’s the data layer you reach through a connector — not the skill, and not the protocol. Separately, the Agent Factory itself is a “source of truth” for how to build AI workers.
Thinking like an architect
When you design an AI worker: give it skills (capability / know-how) and add connectors (to act on the outside world).
Example
Tell Claude “enroll me in Panaversity.” A connector lets it open the site, use your access, add your details, and prompt you to pay. The connector is the technology (MCP) that lets it act beyond its own chat window.
To build capable workers, I need to learn
- How to build a connector, and the design/architecture of MCP — what it hands the worker (tools + resources).
- How to create skills, so a worker uses a tool or MCP effectively.
- Which model(s) fit each job (still an open question — below).
What’s new in Claude
- Isolated sessions — every session runs in its own VM / environment.
- Live artifacts — a session can create editable docs, slides, sheets or designs right there.
- Bidirectional — start from a chat or from a doc / sheet / design; Claude can also start the conversation (it’s proactive), not only the user.
- Share & collaborate — anyone with an artifact’s link can comment on it and work on it.
- Design-first — design something in a design tab, then edit it by hand or with the worker.
- Cowork and Claude Code are now merged.
My note
Prebuilt starts (design, docs, slides) open a progress view and a richer right-hand sidebar; a plain new chat doesn’t show that sidebar. If Claude seems behind on a capability, updating the app / version is the first thing to check.
What “harness” actually means
Correction
The harness isn’t the model(s). It’s the runtime around the model — its system prompt, tools, permissions, memory and the app loop. The same model behaves very differently under different harnesses, and you don’t “merge” a harness: Claude Code is the harness that drives the model here.
Questions I’m still exploring
- Which model or models are best for building vs. running these workers?
- When the “document” is really a CRM or dashboard, how does the worker operate it?
- Good design patterns for pairing a skill with an MCP connector.
Test yourself
recall questions
- Connecting a tool vs. teaching a worker — what’s the difference?
→ A connector gives access; a skill teaches the procedure (the how-to). - What does MCP give an AI worker?
→ Tools (actions it can call) and resources (data it can read), over one open standard. - What is a System of Record?
→ The authoritative data store (source of truth) the worker reads and writes through a connector. - Is the harness the same as the model?
→ No — it’s the prompt + tools + permissions + environment around the model.