Workbenches

Workbenches turn a transformation job or decision cycle into a blueprint your team runs with an agent. Each workbench is a room: the blueprint, the file you are working on, and the conversation together. The blueprint is a set of stages in order, and each stage has a brief — what it reads, what it does, what it writes, and what you check. A person approves every gate before the next stage runs. Status is the files that exist — not a hidden job engine.

This area is separate from Chats. Opening a workbench does not switch you into the chat list.

When it is available

Workbenches stay off until a platform admin shares the feature with your organisation and an org admin enables it. Then Workbenches appears in the rail. The starter workbenches in the Library appear when a platform admin also enables Workbench templates.

Two ways to start

• Use a template when the job already has a shaped blueprint. The Library seeds a starter set — post-merger integration, regulatory resilience, a target operating model, and the architecture jobs (principles, governance, assessment, investment). Use this workbench starts one you can run today; the blueprint is already in place, so there is no factory run first. Without the flag, the Library shows the starter set as coming-soon chrome.
• Start the factory when the job is yours to define. Describe the work your team repeats — the assistant designs a blueprint for it: stages, gates, the way you work. You approve the plan, it becomes a workbench, and you run it. The builder skill is not in Chat or Skills.

Library, factory, and yours

• Library — ready workbenches: the starter templates, or coming-soon jobs before the templates flag is on.
• Factory — the card that designs your own blueprint: Start the factory.
• Yours — the workbench rooms you run: ones you started from a template or designed with the factory.

Every instance is one row with two lives. Start the factory begins life 1: the room opens with the design stages already seeded, and the assistant asks, in the transcript, what repeats, where you pause, and what done looks like. Answer in your own words; it never invents a blueprint from empty answers, and the proposal always waits for your yes before the factory writes it. When the walk test passes and the stages are in place, Use this workbench flips the same instance to life 2 — the workbench. Cards say Factory or Workbench from that life.

Working in a room

A workbench is a room. The blueprint sits on the left, the file you are working on in the middle, and the conversation with your agent on the right; when a stage compiles a visual, it lands on the Stage below. A new factory room is not a blank chat: the assistant asks, in the conversation, what repeats, where you pause, and what done looks like, and waits for your answers. Edit the output file — that is the gate — then run the next stage.

Setup questions are asked once and stored on the instance. Later visits do not re-ask. Connections chosen in Setup are the MCP servers a run can fetch; method skills live in the workbench, not on the composer.

Cycles

Stages show their state from the files: Not started, At the gate, or Passed. The agent's write puts the stage at the gate; reading its output — Mark as read in the file view — passes it. When every stage has passed, the room offers Close this cycle: give it a name, and the outputs move to the workbench's archived runs (runs/) while the tree clears for the next cycle. Closed cycles stay readable on Status, and closing never happens by itself. Deliverables reopen on the Stage: the latest compiled visual stays with the workbench, so a later visit shows it even before a new run starts.

Next: co-create in a chat when the work is open-ended, or keep a visual under Artifacts.