How is this different from an AI coding assistant in my editor?

An assistant helps a developer type. An autonomous coding agent takes a whole issue, runs the stack, tests its own work and opens the pull request. Nobody has to be at the keyboard while it happens.

Where does the agent run?

On a dev seat inside your own AWS account — one isolated machine per tenant, holding your repository, a throwaway database and short-lived credentials.

What if the agent is wrong?

You review a normal pull request with a live preview attached, and you merge or close it. Nothing reaches your main branch on the agent’s say-so.

CARGO MANIFEST № 002 — THE AGENT

Autonomous coding agents that ship from your issue tracker

An autonomous coding agent is only useful if it finishes. TaskShipyard runs one agent per feature, on a machine of its own, and judges it on a single outcome: a tested pull request with a running app behind it.

The issue is the prompt

You tag a GitHub issue with @taskship. That issue — its title, its body, the discussion under it — is the whole brief. There is no chat window to babysit and no prompt to engineer, because the place your team already writes requirements is the place the agent reads them.

If the agent gets genuinely blocked, it comments on the issue and waits. You answer in the issue, like you would for a colleague, and the run continues from there.

Autonomous development means a full environment, not a code completion

Autonomous development stops being a demo the moment an agent has to run what it wrote. Each TaskShipyard run gets a real development environment: the repository checked out on a branch, a local database with your migrations applied, the dev servers up, and cloud infrastructure deployed to an isolated stage in your own AWS account.

That is why the agent can do backend work, database work and UI work in the same run. It is not predicting what the code would do — it is running it.

Verification is part of the job, not an afterthought

Every run has to prove itself in four ways before it is allowed to finish, and the proof goes into the pull request description.

The UI, with screenshots the agent took and read back
The API, with real requests and their status codes and bodies
The cloud, inspected with scoped credentials for that run
The database, queried directly to show the rows the feature writes

What comes back

A branch, a pull request whose description is the agent’s own write-up, the issue updated with the same write-up, and a preview environment still running so a human can click through the result before merging.

Questions we get asked

How is this different from an AI coding assistant in my editor?: An assistant helps a developer type. An autonomous coding agent takes a whole issue, runs the stack, tests its own work and opens the pull request. Nobody has to be at the keyboard while it happens.
Where does the agent run?: On a dev seat inside your own AWS account — one isolated machine per tenant, holding your repository, a throwaway database and short-lived credentials.
What if the agent is wrong?: You review a normal pull request with a live preview attached, and you merge or close it. Nothing reaches your main branch on the agent’s say-so.