Skip to main content

All postsIntroductions

Cloud Agents on TOSE: delegate to AI the way you delegate to a remote teammate

By TOSEPublished Updated 4 min read

Three cloud workstations connected to one desk where a person coordinates the work

TL;DR

Cloud Agents run Claude Code, Codex and Pi in their own environment on TOSE, with persistent disks and a private console. They fit well-defined tasks, several independent tasks in parallel, and orchestration from another AI over MCP. TOSE handles the environment, approved deploy plans, GitHub through a proxy and auto-sleep. The brief, the permissions and the review stay with you.

Narration and captions are in Vietnamese.
Three cloud workstations connected to one desk where a person coordinates the work

Coding agents keep getting more capable. Yet on a laptop, someone still watches the terminal, answers questions and tells the agent to keep going. This article answers a practical question: when should you hand work to an agent the way you would to a remote teammate, and what does TOSE take care of along the way?

What changes with a remote teammate?

With someone sitting next to you, a half-finished explanation is fine and mistakes get fixed on the spot. With a remote teammate, you need:

• a brief clear enough that they can work without constant questions;

• a workspace of their own, with the right tools and just enough access;

• a way to follow progress without checking in every few minutes;

• a result you can verify, such as a pull request.

Cloud Agents on TOSE give Claude Code, Codex and Pi that workspace. Each agent has its own environment and a persistent disk.

The Cloud Agents page with the workspace's agents and the agents available to deploy
Mockup of the Cloud Agents page with sample data.

When should the agent stay on your machine?

Running locally is still the right choice when:

• the task is short and you want to steer each step;

• the task needs data or tools that only exist on your machine;

• you are still exploring and do not yet know the result you want.

A simple rule: work with the agent while the task is fuzzy; delegate once it is clear.

When should you delegate to a Cloud Agent?

The task has a clear brief. You know what may change, which tests must pass and what "done" means.

You have several independent tasks. Each agent runs one task at a time, so three independent tasks can go to three agents, as long as they do not touch the same files or depend on each other's results.

Three agents running three independent tasks at once

You do not want your laptop to carry builds, tests and installs. The agent has its own CPU, RAM and disk on TOSE, and it keeps working after you close your laptop.

You want to orchestrate from another AI. Once you settle on an approach with Claude or ChatGPT, that same AI can hand the task to an agent through the cloud_agent_* tools. See part 3.

What does TOSE take care of?

• A dedicated environment per agent. The /workspace and /home/agent disks persist across stops and restarts. The console has no public address; only workspace owners and admins can open it.

• A plan before every deploy. Each deploy comes with a plan listing resources, the cost ceiling and any blockers, and it runs only after a person approves it.

• GitHub without giving the agent a token. The agent pushes and works with issues and pull requests through the TOSE proxy, with permissions scoped per repository. It cannot merge.

• Task tracking. The agent's Tasks tab shows what is running, what finished and what failed or stopped, with where each task came from and its result.

• Auto-sleep. An idle agent goes to sleep; compute billing stops and files are kept.

TOSE does not replace Claude Code, Codex or Pi. Models run on your own provider API key, and TOSE does not charge for model usage.

What stays with you?

The brief. When an agent works alone, the brief is what keeps it on track. A good brief states what may change, what must not be touched and what evidence shows the work is done.

Permissions. Accepting tasks, enabling push and allowing issue and pull request management are your decisions. A task can read and send back anything the agent can reach, including its sign-in, so grant access only to applications you trust.

Review. "Succeeded" means the agent finished, not that it was right. The diff and the test results are the evidence. Merging is always a human decision.

Next step

Start with one agent, one clear task and one reviewable result, then scale up.

For a product overview, visit the Cloud Agents page.

Frequently asked questions

What are Cloud Agents on TOSE?

Installations of Claude Code, Codex and Pi that run on TOSE. Each agent has its own CPU, RAM, /workspace disk and a private console in the browser.

Does TOSE charge for model usage?

No. Agents use your own provider API key. TOSE bills compute per hour and storage per month.

How many tasks can one agent run at a time?

One task at a time, with no queue. Use several agents to work in parallel.