In Class 31 you gave an agent tools and learned to put approval gates around the consequential ones. Now the question is where that agent lives — inside a product someone else operates, or inside a system you operate yourself. This lesson covers the first option.

The provider supplies almost everything

When you use an agent feature inside ChatGPT, Claude, Gemini, or a similar platform, you are working inside someone else's product. The provider supplies the model, the interface, the execution environment, the limits on how long the agent may run and what it may touch, how memory or chat history behaves, and which tools and connectors are available.

That is the deal. You bring the goal, the instructions, and the files or connections you choose to share. The platform decides what the agent is allowed to reach, how many steps it may take, how long its work persists, and what happens to the data along the way.

This is not a trick. It is the reason hosted agents are fast to start: there is nothing to deploy, no server to maintain, no queue to monitor. You describe the task, attach context, enable the connectors you need, and watch the agent work.

What "connectors" mean here

Each platform offers its own surface for reaching outside services — connected apps, connectors, tool integrations, file access, and similar features. The names and the exact capabilities change often, so treat any specific button label or feature list as a snapshot, not a promise.

What is durable is the idea: a connector is a permission you grant so the model can read from, or sometimes act in, another service through the platform's rules. The platform mediates the connection, enforces its own limits, and can revoke or change it.

Check the current surface before you depend on it:

• ChatGPT connectors and apps: <https://help.openai.com/en/articles/11487775-connectors-in-chatgpt>
• Claude tools, MCP, and connectors: <https://docs.anthropic.com/en/docs/agents-and-tools/mcp>
• Gemini connected apps: <https://support.google.com/gemini/answer/13695044>
• The open protocol behind many of these integrations: <https://modelcontextprotocol.io/introduction>

If a connector you need does not exist, or the platform's rules forbid the action you want, that is your signal that hosted may not fit this particular task.

Where hosted agents fit best

Hosted agents are strongest for work that is exploratory, short-lived, or private to you:

• Exploration. "Summarize these three reports and pull out the disagreements." You supply the files, the agent reads and compares, nothing leaves your account context.
• Drafting. Outlines, first-pass memos, rewrite suggestions, brainstorming against your own notes.
• Private one-off analysis. Upload a spreadsheet, ask for patterns, check the reasoning, export the result. The work ends when the chat ends.
• Personal research. Gather background on a topic, compare sources, build a reading list — with you checking each claim.
• Prototyping an agent workflow. Before building your own agent, run the workflow by hand inside a hosted agent. If the steps, inputs, and outputs work there, you have a specification worth automating later.

The common thread: the stakes of a mistake are low, the work does not need to run while you sleep, and nobody else's production system depends on the output.

What you give up

Control. Specifically:

• Execution limits. The platform caps run length, step count, file size, or connector calls. A task that needs hours of unattended work will hit a wall.
• Memory behavior. The platform decides what persists between sessions and what does not. You cannot redesign it; you can only learn its current behavior and work within it.
• Tool surface. You get the connectors the provider offers, under the provider's rules. You cannot add a custom tool, a private database, or a bespoke approval flow.
• Data boundary. Your inputs pass through the provider's systems under its terms. For personal or public material that is fine. For sensitive customer data, unpublished business data, or anything under a confidentiality obligation, read the terms first and prefer to keep it out.

None of this makes hosted agents bad. It makes them bounded. A capable builder knows the boundary before starting.

A hosted session, concretely

Picture the competitor-summary task running inside a hosted agent. You write the brief ("compare these five blogs, table of claims with links"), upload or link the sources, and enable whatever connectors the task needs. The agent reads, reasons, drafts the table, and asks follow-up questions when a source is ambiguous. You correct it mid-stream, it revises, and you export the result.

Notice what you never did: provision a server, write a scheduler, design a log format, or rotate a credential. That missing work is the speed you are paying for with the limits above. When the task ends, the value ends with it — the session holds no standing machinery for next week's edition. If the summary must repeat weekly, with the same sources and the same checks, you have outgrown the session and should specify the workflow for something sturdier, which Lessons 32.3 and 32.4 cover.

Practical exercise: the control-split table

Pick one real task — for example, "draft a weekly competitor summary from five public blogs."

Draw a two-column table:

Fill in at least three rows per column for your task. Then answer: if the platform changed one thing it controls tomorrow — a limit, a connector, a memory rule — would your task still work?

Finish line: one completed control-split table plus a one-sentence verdict ("hosted fits" or "hosted does not fit, because…").

Verify it: point to the platform's current docs page for one connector or limit you listed. If you cannot find it documented, mark that row as "assumed — recheck."

Common failure mode: listing what the agent *can* do while ignoring who decides. "It can browse" is a capability. "The platform decides which sites, how many pages, and for how long" is the control split that matters.

Check your understanding

1. Name four things the provider supplies in a hosted agent environment. 2. Why is prototyping inside a hosted agent a good step before building your own? 3. Give one example of a task where the platform's data boundary alone rules hosted out.

Next, we look at one hosted capability in detail: agents that browse and operate the web on your behalf — what they do well, and where they break.