Agents

Workbench uses Claude or Codex as the main tutor for a playground session, with reviewer and support agents available when the work needs another pass.

Main tutor agent

The main agent is Claude or Codex running inside Workbench.

In a normal coding setup, that agent is an open-ended assistant. In Workbench, it becomes a tutor: it reads the playground step, explains what the student is building, helps with code, checks the result, and moves the session forward only when the work makes sense.

Workbench keeps that tutor attached to the playground step, visible code changes, and local runtime signals.

Reviewer agents

A reviewer agent is useful when the work needs a second pass.

Use it to inspect a completed step, check whether the implementation matches the task, look for missed edge cases, or review a failing runtime signal before the main session continues.

Support agents

A support agent can help with bounded side work while the main session stays focused.

Examples include investigating a log message, checking a small part of the repository, or summarizing a local failure mode before the main agent applies a fix.

When to use another agent

Use another agent when the task benefits from a separate review or a narrow investigation.

Keep the tutor agent responsible for the active learning path. That keeps the session understandable and avoids turning the workflow into several unrelated conversations.

If you have questions, visit our Discord