Coding Agents and Agent Studio

Agent Studio is the workspace catalog for Standard and Coding agents. Workspace administrators create agents, validate their dependencies, test them privately, and review their runs in one place.

  • Standard agents work inside Windshift with the tools and access granted to their profile.
  • Coding agents work in connected repositories through an authorized runner. They can prepare changes, push a branch, and open a draft pull request.

Existing coding-agent bindings continue to work as legacy profiles. You do not need to convert them when you upgrade to 0.8.5.

Open Agent Studio

Open a workspace and select Agents. The catalog shows each specialist's type, runtime, identity, availability, and most recent run.

Only workspace administrators can create or change agents. Other workspace members can view agents that their permissions allow and use agents that are Ready.

Create an agent

  1. Open Agents in the workspace.
  2. Select Create agent.
  3. Choose an approved specialist template.
  4. Select Standard or Coding when the template permits it.
  5. Set the agent name, handle, purpose, identity, instructions, and model connection.
  6. For a Standard agent, select the optional tool and access groups.
  7. For a Coding agent, select the repository, base branch, runner pool, and runner setup.
  8. Save the agent.

A new agent starts as Draft. Agent Studio preserves an unfinished creation form in the browser, but it never stores a plaintext runner registration token after the setup view closes.

Configure the profile

An agent profile has these tabs:

  • Overview: Name, handle, purpose, identity, model, repository, and runner configuration.
  • Instructions: The specialist instructions added to the runtime instructions.
  • Knowledge: Reusable workspace knowledge packs attached to the profile.
  • Tools and access: Required and optional capability groups available to a Standard agent.
  • Test: A private test prompt and its live run output.
  • Runs: Recent production and test runs for this profile.

Saving profile details, instructions, knowledge, tools, a model, a repository, or a runner creates a new Draft version. Check readiness again before you make that version Ready.

Check readiness

The Readiness panel validates the live dependencies for the current profile. Depending on the agent type, it checks the model connection, identity, permissions, repository, and runner.

When every required dependency is available, select Make Ready. A Ready agent can receive normal work. If you archive an agent, it becomes unavailable. Restoring an archived agent returns it to Draft and runs another readiness check.

Readiness is structural. It does not prove that an agent will produce a useful answer for every prompt.

Run a private test

Workspace administrators can use the Test tab before they make an agent available to the team.

  1. Enter a test prompt.
  2. Start the private test.
  3. Watch the run output.
  4. Cancel the run if it no longer needs to continue.
  5. Review the result and update the profile when necessary.

Private testing is optional and restricted to workspace administrators. A passing private test does not change the readiness result.

Start normal work

You can start a Coding agent run from a work item in two ways:

  • Assign the work item to the agent.
  • @mention the agent in a comment without changing the assignee.

Mentioning an agent that is already working on the item does not start a duplicate run.

A Coding agent checks out the authorized repository, makes changes in an isolated container, and pushes its result branch. A run without commits does not push an empty branch or open an empty pull request.

The work item and the agent's Runs tab show run progress and history. Availability labels show whether an agent is Ready, Draft, offline, paused, invalid, or archived. An offline Coding agent can queue work until its runner returns.

Runner security

Coding agents run in fresh containers. Git and model credentials stay on the Windshift server. The runner and agent receive short-lived, scoped access for the current run.

The runner controls the Docker daemon on its host. Prefer a dedicated runner host for untrusted code. See Coding Agent Runner for installation and network guidance.

Manage agent templates

System administrators manage template overrides under Admin > AI Connections > Agent Templates. An override can change a template's name, default profile type, instructions, and enabled state.

When no override exists, Agent Studio uses the built-in template. Disabling or changing a template affects new profile creation. It does not rewrite an existing agent profile.