What is a cloud coding agent?

Quick answer

A cloud coding agent is an AI agent that reads, changes, tests and ships code from a remote machine instead of your laptop. You give it a task in chat, a ticket or a schedule; it clones the repository into an isolated sandbox, does the work, and returns a pull request. Because it runs in the cloud, it keeps working when your laptop is closed and anyone on the team can start or steer it.

How a cloud coding agent works

Every cloud coding agent has the same four parts. The names differ by vendor, but the parts do not.

• A harness. The program that runs the model in a loop: read files, run commands, edit code, check the result. Claude Code, Codex and Gemini CLI are harnesses. The harness decides how the agent plans, edits and checks its work; the model is only one input.
• A sandbox. An isolated machine with your repository, toolchain and credentials. The agent can install packages and run tests there without touching anyone's laptop.
• A way in. Something that starts the work: a Slack mention, a web chat, a ticket, a cron schedule or a webhook.
• A control plane. The service that decides who may start a session, which credentials it gets, where it runs and what is kept afterwards. See what an agent control plane is.

A local agent has only the first part. Everything else is your laptop and you. That is fine for one engineer at a desk. It stops working when the task should run overnight, when a product manager wants to ask for a fix, or when ten engineers each configure the agent differently.

What a session looks like, start to finish

Here is one request as it runs on Nori. The steps are the same on most platforms; what changes is how much you have to wire up yourself.

1. An engineer writes in a Slack thread: @nori the CSV export drops rows with a null region. Fix it and add a test.
2. Nori starts a sandbox with the team's repositories already cloned, the team's skillset installed, and the credentials that engineer is allowed to use.
3. The agent reproduces the bug, writes a failing test, fixes the export, and runs the test suite.
4. It pushes a branch, opens a pull request, and replies in the thread with the link and a two-line summary.
5. A reviewer comments on the pull request. The agent picks up the comment, pushes a follow-up commit, and replies again.
6. The session is released. The conversation, the branch and anything the agent saved to shared memory remain for the next session.

Nobody opened a terminal, and nobody's laptop had to stay awake.

Cloud coding agent vs local coding agent

The two are not rivals. Most teams keep a local agent for pair-style work at the keyboard and use cloud agents for everything that can run without someone watching. The longer comparison is in cloud coding agent vs running Claude Code locally.

What to check before you pick one

Vendors use the same words for different things, so ask about the parts, not the label.

• Harness choice. Does it run one vendor's agent, or the harness your team already uses? Nori ships its own agent, Jiro, and also runs Claude Code, Codex, Gemini CLI and any ACP agent.
• Who pays for tokens. Some platforms bill model usage through their own credits. Nori lets you use your existing Claude or ChatGPT subscription, API keys or LLM gateway and never charges for tokens. See do cloud coding agents mark up token costs.
• Ways in. Chat, tickets, schedules and webhooks should all start the same kind of session, so automations and people share one history.
• Shared context. Instructions, skills and memories should live at the team level so every session starts with how your team ships.
• Where the control plane runs. If your security team needs everything inside your own cloud, ask whether the control plane moves too, or only the runners. Nori deploys the full stack into your AWS VPC or Azure VNet.
• Identity and credentials. Sessions should act with scoped credentials, tied to your identity provider, not with one shared admin token.

For a side-by-side of the main products, see cloud coding agents compared.

Where cloud coding agents fit best

The work that suits a cloud agent is well-defined, testable and repeatable. Teams usually start with three kinds.

• Requests from chat. Small fixes, questions about the codebase, and follow-ups on pull request reviews, asked where the team already talks.
• Scheduled work. Dependency updates, flaky-test triage, documentation checks and data health checks that run on a cron schedule and post a summary.
• Event-driven work. An error-tracker webhook that becomes a reviewed fix pull request, or a ticket assignment that becomes a branch.

The same pattern extends past application code. Security teams run continuous red-team and blue-team loops, and data teams run governance checks on every new table, using the same sessions, schedules and review steps.

Frequently asked questions