Agents
Agents is the Workbench view for monitoring several agent sessions at once. It replaces the old assumption that Grid is Wardian's global page: Grid is now one display mode inside this view.
Use Agents when you want live terminals or normalized chat activity arranged together, then switch to a focused presentation when one agent needs more space.
For a visual map of an agent's session, workspace, class, and skills, see Key Concepts.
Open Agents
Press Ctrl+P / Cmd+P, select a pane's + button, or use the empty-pane Home state, then choose Agents. Wardian keeps a single Agents surface and focuses it when you open it again.
The right roster controls which agents auxiliary tools target. It does not filter Agents automatically. Use Filter agents in the surface to narrow the visible cards by agent name, description, class, provider, or workspace.
Remember an Agent's Purpose
Select one agent, open Agent Configuration from the left rail, and add an optional Description. Use it as a short memo about that specific agent's responsibility, such as Owns frontend release follow-up.
Wardian keeps the memo available in agent configuration and automation assignment choices, while the compact roster shows only the agent's class. Roster and assignment searches include the memo. Descriptions do not change the agent's instructions, skills, permissions, or provider session, and saving one does not restart the agent. New agents start without a description, so the spawn automation has no additional required step.
After restarting Wardian, agents can appear in the roster while their sessions are still restoring. Configuration saves and lifecycle actions for such an agent wait for its restoration to finish. Once a save succeeds, later resume actions use the saved settings, including the choice to start a fresh provider conversation. This also applies to paused agents.
Interactive terminals use the same color capabilities when Wardian is opened manually or by automation. The first visible terminal fits an unowned session to its pane automatically, including when the session finishes restoring after the view opens. A second view of an already owned session remains a mirror of that owner's terminal geometry.
Open an Agent from Another View
When Graph or Inbox uses Open Agent, Wardian first targets an existing Agents view anywhere in the Workbench. It can activate an inactive Agents tab in the same pane or another pane, select the requested card, and scroll it into view. If the target is hidden by a zoomed pane, Wardian restores the split layout before moving focus there.
If the view's local agent search or status filter would hide the requested card, Wardian clears only that blocking filter before revealing it. Filters that already include the requested agent stay in place.
In Garden, double-click or Enter opens the agent's composition. Its explicit Open agent session action opens or focuses the canonical Agent Session terminal through contextual Workbench navigation, preserving Garden's breadcrumb and camera; it does not reuse the Agents view.
The roster's explicit Open and Open to Side controls remain terminal actions: they open or focus Agent Session tabs instead of changing the contextual Agents destination.
Choose a Layout Mode
The surface header provides three modes:
- Auto always keeps a multi-agent roster visible. It uses multiple columns whenever the pane can fit at least two useful cards side by side, and stacks the full roster into one scrolling column when it cannot. With one visible agent, that card naturally fills the surface.
- Grid keeps multiple visible cards and enables direct grid sizing and ordering controls. The layout can scroll when the available area is smaller than the cards' usable floor.
- Single shows the focused agent as one full-surface card. Use a card's Maximize control to focus it. Minimize restores the last explicit multi-agent mode, Auto or Grid, and reveals the roster again.
Auto derives its layout from the current surface dimensions and targets comfortable working cards before adding rows or columns. Terminal cards prefer about 640 x 450px and Chat cards prefer about 480 x 450px; smaller hard floors are reserved for genuinely constrained panes. Extra rows scroll. Auto does not overwrite the manual Grid arrangement, so you can return to Grid without losing the layout you tuned.
Terminal and Chat Cards
Each card can show either:
- Terminal for the provider's real PTY, approvals, raw keybindings, and interactive TUI.
- Chat for normalized messages, tool activity, approval choices, and a compact prompt composer.
Chat messages and activity rows share a centered, readable transcript column that expands fluidly on narrow surfaces.
Use the Terminal / Chat button in the card header for a temporary per-agent override. The default comes from Settings > Agents > Agent card display. Unsent Chat text stays with that agent when you switch modes.
A compact Subagents badge appears in the card header only when Wardian has verified one or more currently active provider-spawned workers through that agent's provider ancestry. The visible badge shows the current active count; the watchlist uses an icon and that count to keep the roster compact. Requested, running, and waiting workers remain active, and waiting workers show an attention marker. Hover the badge for the full summary or select it to inspect active workers and their evidence. Terminal and unknown records remain available to automation run history and retention, but they do not create a root badge, root warning, or root list entry. The inspector preserves evidence coverage, available capabilities, own usage, and descendant combined usage without presenting retained runs as active. Observe only identifies providers for which no canonical follow-up, resume, or interruption control has been verified. The badge reports attention without creating another agent card or watchlist row. Inspect automation workers from their Monitor run and Observe node; a headless worker does not receive a fictional terminal. Older backends without positive active-state evidence do not create a root badge.
In Terminal mode, click inside the terminal before typing. When the same agent terminal is visible elsewhere, the clicked presentation explicitly requests ownership; merely tabbing through the UI does not steal it. A Mirror remains read-only until ownership transfers. Reclaimed renderers restore and fit automatically when visible.
During initial display, new and resumed terminal frames stay hidden while their geometry settles. A visible, mounted renderer retries the fit automatically, so temporary sizing instability does not require a window resize to reveal it. If restoring a reclaimed renderer fails, its recovery notice provides a Retry action.
Arrange and Focus Agents
- Drag card headers to reorder agents in Grid mode.
- Drag row and column gutters to adjust manual Grid sizing.
- Use Maximize on a card to enter Single mode for that agent.
- Use Minimize to leave explicit focus and return to the previous Auto or Grid mode.
- Right-click the background and choose Reset Grid Layout to discard the manual Grid arrangement without resetting the whole Workbench.
- Use the card menu for agent lifecycle actions. These actions affect the runtime; closing the Agents tab does not.
Auto adds columns only when the pane can support their preferred working width and permits vertical overflow. Below that threshold it switches to a one-column roster without hiding agents. Grid always honors the user's chosen tracks and row sizing instead of silently changing the user's arrangement.
Important Boundaries
- Closing Agents closes only that Workbench presentation. Agents keep running.
- Closing an individual agent-session tab also leaves the runtime alive. Use Pause, Start Session, Restart Session, New Session, or Delete Agent only when you intend to alter the agent lifecycle.
- Roster selection and Agents focus are separate. Roster selection targets tools; Agents focus chooses the agent shown in Single mode.
- Grid is not a global navigation destination. It is an Agents mode retained for multi-agent compatibility and control.
Configure Provider Launch Options
Open an agent's configuration and expand Advanced Settings to override options for that provider. Wardian shows only options consumed by the selected provider. Switching providers replaces the provider-specific configuration so Claude, Gemini, Codex, Antigravity, OpenCode, and Pi flags do not leak into one another.
Options marked Headless apply to background deliveries and automation runs that use the agent profile. They do not alter the provider's interactive terminal. Wardian owns structured output flags for those runs so provider JSON or stream formats cannot break delivery parsing.
Codex sandbox and approval values inherit the provider-wide defaults under Settings > Agent Runtime > Codex unless the agent has an explicit override. Gemini's sandbox option requires Docker or Podman on the machine running the provider CLI. If neither is available, Gemini rejects the launch before it can begin a session. Pi's project approval options control project-local configuration, extensions, and skills. They do not sandbox Pi's tools or extensions. Use Custom Arguments for a current provider flag that Wardian does not model; custom arguments are passed only to the selected provider.