Skip to main content
The workbench is a full-screen terminal UI where you build agents, plans and tools by talking to graph’s built-in agents, then run and debug them in the same window.

Open it

It opens at the front desk, which asks what you want to build. To go straight to a mode, name it, and optionally the file to open:
graph wb is short for graph workbench. The workbench needs an interactive terminal. In scripts and CI, use graph plan run.

Layout

The workbench with the chat pane on the left, where plan_editor describes unreleased_changes, and the workspace on the right, showing the plan's steps and the detail of the selected row.

unreleased_changes open in the workbench

  • Chat, on the left: you talk to the workbench’s agents. The pane title names the agent you’re talking to.
  • Workspace, on the right, with three tabs:
    • 1 the plan, agent or tool you’re working on, with the selected row’s detail below it.
    • 2 the tool catalog the agents draft from.
    • 3 the latest run.

Who you’re talking to

The front desk hands you to the plan drafter, the agent drafter or the tool drafter. Once a plan is drafted or opened, the plan editor takes over. The plan drafter and plan editor can hand you to the tool or agent drafter when a plan needs a new tool or agent, and those drafters hand you back when they’re done. To switch by hand, type /agent <name> in the chat. See Built-in agents for what each one does.

Saving

Ctrl+S saves what’s in the workspace: plans to ./.graph/plans/, agents to ./.graph/agents/ and tools to ./.graph/tools/, ready to commit. A file you opened is saved in place.

Keys

Tab moves focus between chat and the workspace. Most keys act on the workspace, so press Tab first when you’re typing in chat. While a debug run is paused:

Mouse

Logs

The workbench writes its log to workbench.log in the data directory, not to the terminal. Set GRAPH_WORKBENCH_LOG or [workbench] log_path to change the path, and GRAPH_LOG to change what it records.

Guides

Draft a plan in the workbench

Describe a goal and get a plan you can run.

Debug a plan

Pause at a breakpoint, find the bad input and fix the step.

Create an agent in the workbench

Draft an agent with tools and save it.

Create a tool in the workbench

Draft a tool, test-run it and save it.