Coding Agents

Coding agents are AI agents that pick up a work item, do the work in your repository, and open a pull request with the result. Where the AI Features assist you while you work, a coding agent takes a whole task off your plate: it checks out the repo, makes changes, pushes a branch, and opens a draft pull request linked back to the item.

Agents run in isolated, throwaway containers, and your Git and model credentials never leave the Windshift server. Self-hosting operators set up where agents run. For that side, see Coding Agent Runner.

Setting up an agent

An agent is configured per workspace as an agent binding, in the workspace settings under Agent bindings. A binding wires together:

  • An agent identity (a service user that appears as the assignee and as the author of the branch and pull request).
  • A repository, through a Git connection on the workspace.
  • An LLM connection the agent uses to do the work.

Once a binding exists, the agent shows up like any other member of the workspace and can be assigned work.

Starting a run

There are two ways to start a run on a work item:

  • Assign the item to the agent. Assigning a work item to a bound agent user starts a run on that item.
  • @mention the agent in a comment. Mentioning a bound agent in a comment starts a run on that item without changing the assignee or status. This is a lighter touch when you just want the agent to take a look. Mentioning an agent that is already working on the item does not start a second run.

When a run produces changes, the agent pushes a branch and opens a draft pull request. A run that produces no commits does not push an empty branch or open an empty pull request.

Watching a run

  • The work item gains an Agent log tab that streams the run's progress live as it happens, so you can follow what the agent is doing.
  • Assignment and mention pickers show whether an agent is currently available to take work, so you are not handing an item to a busy or offline agent.

Shaping an agent

Each binding can be tailored without changing its identity or repository:

  • Custom instructions. Give the agent a persona or extra ground rules (for example, "You are our release manager"). These are appended to Windshift's standard operating rules, never instead of them.
  • Skills library. A workspace can maintain a library of markdown knowledge packs and attach them to an agent. The agent sees a short index of its skills and fetches the full text of a skill only when it is relevant, so a large library does not slow runs down. Skills are managed in the workspace settings alongside bindings.

Both custom instructions and skills are edited on the binding under Persona & skills.