Skip to content
The AI Agent Factory

The Architecture

Approved by
Panaversity ·
10 min read

How the new apps work behind the screen, with the evidence

Where this section fits

This is the fourth section of Claude and ChatGPT 2.0: the technical edition. It is optional. Read it for the evidence and the engineering, or come back when you build AI Workers for a business. The first section introduced AI Workers, the borrowed desk and Brightline's Friday payment run. Tour of the Apps documented each feature. Behind the Screen explained the same ideas with one analogy: the model is a consultant, the harness is a project manager, and the sandbox is the borrowed desk.

This section adds the evidence: what was observed in a real Claude session, what each vendor documents, why both chose the same design, and what a business must add. It does not repeat the Tour's feature lists.

What the architecture is

In the desktop app or on the web, a Claude session runs in the cloud by default. A session is one running conversation or task. A request travels this path: UI → Anthropic backend → harness → model on Anthropic's inference servers. The UI is the app or web page you type into. The backend is the servers behind the app. Inference servers run the model. This section calls the UI, the backend, the harness and the model its four layers.

The firsthand observations come from one cloud Claude session that the Panaversity team inspected on November 8 and 9, 2026. The appendix gives the method.1 ChatGPT also reported an inspection of its own code environment. It did not show ChatGPT's harness or inference.2 The rest of the OpenAI side rests on OpenAI's documentation.

  • [Vendor-documented] Cloud sessions are the default. The agent loop and code execution run on Anthropic's servers. Each session gets its own sandbox, "created when the session starts and destroyed when it ends".3
  • [Observed] In the session examined, the harness was Claude Code. It ran inside a Firecracker microVM, a small virtual machine that starts very quickly. The session had been started from a claude.ai chat.
  • [Inferred] The model does not run inside the sandbox. The sandbox is where the model's tool calls run and where its files are kept.
  • [Vendor-documented] Each scheduled task run starts a new session. The exception is a run inside a project, which can continue an existing thread.4,5
  • [Inferred] No VM runs permanently for an account. None is documented, and none was observed.
  • [Inferred] What changes from one surface (the web, the desktop app or a terminal) to another is where the harness runs: in a cloud sandbox, on your own device, or on a server you manage. In Claude's own apps, on Anthropic-hosted environments, the model runs on Anthropic's servers by default.

How this section is organized

Why both vendors chose the same design comes first. It explains the goals.

The first half shows how the runtime works, with Claude as the observed example: request flow, the sandbox and its state, cloud environments, scheduled tasks, projects, where to run work and security. It ends with a worked example and ChatGPT and Codex.

The second half shows how these runtimes become AI Workers: are they AI Worker platforms?, from personal to business worker and what comes next. Three appendices hold the reproduction steps, business-layer details and runtime details.

Note

Words used in this section

  • Model: the stateless inference service that reads a context and returns text or a tool call. Stateless means it keeps nothing from one call to the next. Inference means running the model to produce an answer. Context is everything the model reads in one call.
  • Harness: the agent loop around the model. It builds the context, runs tools and tracks progress. Claude Code and Codex are harnesses. It can run inside the sandbox, as observed here, or apart from it.
  • Sandbox: the isolated place where tool calls run, the borrowed desk. It can be a cloud VM or container, or a limited space on your own computer.
  • Virtual machine (VM): a complete computer, simulated in software. A container is a lighter isolated space that shares the operating system of the computer it runs on.
  • Session: one running conversation or task, with its own sandbox and history.
  • Runtime: the harness plus the sandbox, as a product you run work on. For Claude, it is the Claude app, Claude Tag, Managed Agents or the Agent SDK (with the Agent SDK, you host both). For OpenAI, it is ChatGPT Work and Codex.
  • Scheduled task: a saved prompt that starts a new session on a schedule or when something happens (a trigger). Both vendors offer them.
  • Worker principal: the identity a worker acts under. It is separate from the people who sponsor or approve the worker.
  • Business layer: what an organization adds around the runtime: identity, durable records, enforced authority and audit.

Note

How to read the evidence labels

Each claim in this section carries a label in square brackets. The label says how we know it.

  • Observed: seen directly in the Claude session examined: command output, running programs, the environment and behavior.
  • Reported by ChatGPT: results that ChatGPT reported from commands it ran in its own tools, at our request. Nobody else has reviewed the raw output.
  • Tool-documented: stated in the descriptions of the tools available inside the session, but not seen in action.
  • Vendor-documented: stated in Anthropic's or OpenAI's public documentation, with the source in a footnote.
  • Inferred: our conclusion from the observations and documents. The vendor does not state it.
  • Unknown: not visible from inside a session, and not documented.

Products change often, and features differ by plan. This section describes them as of November 9, 2026. If your app looks different, trust the vendor's help page. The architecture still holds.

Check yourself

Question 1 / 3

0 answered

In this section, what does the label Inferred mean?

Sources

  1. The observations come from one cloud Claude chat session on November 8 and 9, 2026, and may not hold for every plan, account or surface. The commands and their limits are in the appendix.↑
  2. At our request, ChatGPT ran commands in its own Python and terminal tools, in a separate ChatGPT conversation, and reported the results. The raw output was not independently reviewed.↑
  3. As verified on November 9, 2026. Claude Cowork architecture overview, Anthropic.↑
Owner
Panaversity

On this page