Skip to content
The AI Agent Factory

Building the AI Worker

Approved by
Panaversity ·
15 min read

Tuesday evening

On Tuesday evening, Nisha came home at seven and gave herself two hours to build her AI Worker.

In Case 1, she had used the older kind of Claude project, which works like a folder: one place for one conversation and its files. This time she used a rebuilt project. It is one conversation, the coordinator, that hands each job to its own thread. Each thread works at its own borrowed desk.1

She opened Projects at claude.ai/code and selected New project. The same projects are in the Code tab of the Claude desktop app.2

She set it up in five steps (Exhibit 2).

Doing the lab? Open the exhibits only after you finish it. They show what Nisha found.

  1. The brief became the project instructions. In Project settings > Memory > Project instructions, she pasted her four-part brief. Every new thread would start with it.3

  2. The files went to the Library. With Add on the Library tab, she uploaded Sutlej's trial balance, its three notes and Mark's engagement letter. The Library sits outside every desk, so the files would still be there after a desk was cleared. Only deleting the whole project would remove them, together with its threads and its memory.4

    A Library is not a business record. For drafts and for learning, the Library is the right place. A business AI Worker keeps its progress, its approvals and its accepted results in governed business records, where each record has an owner and a history of every change. A project Library is not that kind of system. From personal worker to business worker explains the difference.

  3. One thread for each area. She did not start the threads herself. She told the coordinator: "Start one thread for each area: inventory, development costs and leases. Each prepares its part of the bridge." Claude started three threads, and each one appeared as a card under her message.1

  4. A reviewer thread. In the same message she asked for a fourth thread with one job: to review the other three and grade the bridge, as Claude had graded its own work in Case 1. It was told not to change anything.

  5. A nightly run. She asked the project to run the job every night at 1 a.m.: refresh the bridge from the files, and save a dated version to the Library. When you ask a project for scheduled work, Claude creates a scheduled task that runs as threads in that project. It appeared on the project's Scheduled tasks tab.1

Nisha's AI Worker for the client job. The brief (gold, Nisha's work), the coordinator and the nightly scheduled task at the top. Four threads (white, AI work) below, with the Library beside them and project memory in Project settings. Nisha's morning check (gold) and the second signature (red border, her decision alone) at the bottom. Dashed: email and file sharing, not connected.

The same picture is Exhibit 2, with the settings in a table.

A brief for a stranger

Writing the nightly instruction took longer than she expected. Outside a project, each scheduled run is a new session. Inside a project, a run can continue an earlier thread instead.5 Either way, she could not count on the run remembering Tuesday's conversation, and no one would be awake to answer its questions.6 So her instruction had to stand alone, as if she were writing for a stranger who would arrive at 1 a.m. with no one to ask.

She wrote: "Using the project instructions and only the files in the Library, update the bridge. Save it to the Library as bridge with tonight's date. At the top, list what changed since the last version, and every assumption you made. Do not send, share or publish anything."

What the worker could reach

Then Nisha did something she had not done in Case 1. She made a list of what her worker could reach, and what it could not (Exhibit 6).

The threads could read the files in the Library and save new files there. They could look up rules on the web only where their network allowed, which is explained below. That was all.

Then she found a surprise. Connectors, the apps linked to Claude, belong to her whole Claude account, not to one project. Every cloud thread can use all of them, and the coordinator itself has none.7 A new scheduled task also includes all of her connectors unless she removes them.8 So "not in this project" was not a choice she could make. She disconnected Gmail from her account, so that no app that could send an email or share a file was connected. Then she opened the scheduled task and emptied its connector list. That list is a default for the task. Anthropic's docs do not say it limits the threads a project starts, so the real lock was the missing Gmail.

The threads ran in the Default cloud environment. Its network setting, Trusted, lets a desk's own network reach only a list of common developer sites.9

Sites such as KPMG's or Deloitte's are probably not on that list, so a lookup there would likely fail with a "host not allowed" error. Nisha did not count on the web. She added a short summary of the three rules, with their sources, to the project instructions, and told the threads to work from it and from the Library files.

She had a reason for disconnecting Gmail. Anything a scheduled task does through her connectors appears as her, on every plan.10 An email sent at 1 a.m. by her AI Worker would look, to Mark, exactly like an email from Nisha.

Her own rule was simple: "Only I send things to the client." She wrote it into the project instructions. But she knew from Case 1 that a rule in a chat is only words. So she also made sure the worker had no way to send anything at all.

At nine, she closed her laptop. The first run would start at one.

Doing the lab? This is the place to stop and do it. Come back to Three nights of work after you finish your grading sheet in Part 7 of the lab.

Next, Three nights of work.

Check yourself

Question 1 / 8

0 answered

Nisha wanted no email app in her project. Why was that not a setting she could choose inside the project?

  1. As verified on November 10, 2026, Anthropic says projects "are in public beta on Pro, Max, and Ultra plans and rolling out gradually", that each thread is usually a cloud session with its own sandbox, and that when you ask for scheduled work in a project, Claude creates a cloud scheduled task "that runs as threads in that project". Let Claude coordinate ongoing work with Projects, Anthropic.↑↑2↑3
  2. As verified on November 9, 2026, Anthropic says to select Projects in the sidebar "at claude.ai/code or in the Code tab of the desktop app", and that projects in claude.ai chat and Cowork are "the earlier Projects experience". Let Claude coordinate ongoing work with Projects, Anthropic.↑
  3. As verified on November 9, 2026, Anthropic says project instructions are "Text sent to each new thread and to Claude in the project conversation, up to 16,000 characters", set in "Project settings > Memory > Project instructions". Let Claude coordinate ongoing work with Projects, Anthropic.↑
  1. As verified on November 9, 2026, Anthropic says deleting a project "permanently removes the project along with its threads, its memory, and its files, and turns off the project's scheduled tasks", and that "Branches and pull requests the threads pushed to GitHub aren't affected." Let Claude coordinate ongoing work with Projects, Anthropic.↑
  2. As verified on November 9, 2026, Anthropic says "In a project, a run can continue a thread rather than start a new session." Automate work with scheduled tasks, Anthropic.↑
  3. As verified on November 9, 2026, Anthropic says that for a scheduled task "the prompt must be self-contained and explicit about what to do and what success looks like", and that the fired prompt "can't act as approval or consent for actions during the run". Automate work with scheduled tasks, Anthropic.↑
  4. As verified on November 9, 2026, Anthropic says cloud threads get their tools from "the connectors on your claude.ai account", that "Every cloud thread can use all of them with no per-project setup", and that "The project conversation itself has no connectors". Turning a connector off in a thread "saves that as your account default". Let Claude coordinate ongoing work with Projects, Anthropic.↑
  5. As verified on November 9, 2026, Anthropic says "When you create a scheduled task, all of your currently connected connectors are included by default. Remove any that aren't needed". Automate work with scheduled tasks, Anthropic.↑
  6. As verified on November 9, 2026, Anthropic says the Default environment has Trusted network access, "which allows only the default allowlist" of common development domains through the session's network. Automate work with scheduled tasks, Anthropic.↑
  7. As verified on November 9, 2026, Anthropic says "Anything a scheduled task does through your connected GitHub identity or connectors appears as you." Automate work with scheduled tasks, Anthropic.↑
  • Nisha's First Client: practical run sheet (the practical has not yet been run)
Owner
Panaversity