AgentDeck
AgentDeck is the working surface for driving coding agents. A workspace points at a directory and picks a harness; a session runs inside it; the workbench is where you read and steer that session.
Everything here renders through one session layer, so the same conversation, the same permission prompt and the same queue serve Claude Code, Codex, OpenCode, Qwen Code, Hermes, maki and mechabedo's own agent alike.
Cards that only you answer
Most cards a session puts in the conversation are questions, and an orchestrator job you told to answer for you may answer them. A consent card is different: it is marked consent, says Only you can answer this, and no job or agent can settle it. A session raises one to add a permission rule, to arm a supervising agent, to run a command that prints a key, or to open a vApp to runners and webhooks. It names the action and everything that would move, and the session learns only your yes or no. On the phone it opens the session rather than offering a button on the notification.
A key card is a consent card with a masked field. What you type goes straight to the setting it names, or to a one-time handle for the tool that asked, and never into the conversation, the transcript or the agent's context; the answered card says Provided. Have an agent get it opens a second session that works out the command that prints the key, and shows you that command on a card before it runs.
A session also knows who launched it. One a member of your fleet started is offered only the tools a member may use, and one another installation opened is offered no terminal, window or vApp tool.
How the pieces fit
Drawing the diagram…
The pages
| Page | Contents |
|---|---|
| Workspaces and sessions | Creating one, picking a harness, what a session is, and what survives a restart |
| The workbench | The three columns, the prompt line, the five tabs of the right pane, and the board: the agent royale, every session an agent in its own office |
| Workflows | A staged run armed on the prompt line: interview, refine, dispatch, observe, conclude |
| Orchestrator | Jobs that supervise a session from outside it |
| Transcripts | The durable record of every run |
| Techniques | Ways of working the pieces above make possible |
Related pages outside this section: Harnesses for what each agent is and how it is installed, Loadouts and kits for what a launch decided, Enhance for what a session carries, and The window for the editors, terminals and bridges beside the deck.