Using the Dashboard¶

The AgentWeave Hub includes a web dashboard at http://localhost:8000 that gives you real-time visibility into every registered project and its agents.

The left rail is a live collection of registered projects and their agents. Selecting a project opens project-scoped content tabs: Overview, Tasks, Spec, Jobs, Activity, and Environment. The browser URL records the selected project, view, agent, and AgentWeave conversation, so reload and back/forward navigation restore the same destination.

Use the project controls in the rail to open an existing directory or explicitly create and register a new one. A missing or moved project remains visible; use Locate in Environment settings to rebind it without losing history. In Docker, submitted paths must be container-visible below the configured mounted workspace root.

Dashboard Sections¶

Overview¶

The main landing page showing:

Agent health grid with status and context usage
Task summary with status counts
Recent activity ticker
Unanswered questions interrupt card
Quick navigation to any section

Tasks Board¶

All tasks with status, priority, assignee, requirements, acceptance criteria, and deliverables. Click any card to expand and see full details.

Tasks are organized in a Kanban-style board by status: - Pending → Assigned → In Progress → Under Review → Completed → Approved → Revision Needed - Rejected tasks are shown in a collapsible section

Messages Feed¶

Inter-agent messages with expand-to-read for long content. The message type and linked task are shown inline.

Human Questions¶

Questions agents have asked you. Answer directly in the dashboard without switching to your terminal.

Activity¶

Live event stream and per-agent output log. See what each agent is doing in real time.

The Agent Activity panel includes: - Activity Tab: Live event stream from all agents - Output Tab: Real-time console output from agent sessions - Info Tab: Agent configuration, roles, and runner type

Agent Chat¶

Each agent owns durable conversations, and you switch between them in the left rail rather than from a selector inside the chat — the rail lists every agent's conversations, and each one is a place you can return to.

Conversation list — grouped by agent, or flattened into a recency view for scanning across agents. A conversation waiting for you is marked on its row, so you can see it without opening it.
Timeline — operator input, the agent's own output, and agent-to-agent traffic in both directions, each placed by its recorded association rather than inferred from timestamps.
Composer — sets the model, the reasoning effort and the permission posture for the next message, and shows how full this conversation's context is. What each posture does and does not check is in Permission postures.
Checkpoint — writes a record of where the work stands so it can continue in a fresh conversation, rather than letting one thread grow until it degrades.

A conversation records where it came from: one you typed, one another agent started, or one a scheduled job or loop firing created.

The side panel¶

A conversation can open a panel beside it, hosting one tab at a time. The plus affordance offers whatever is not already open:

Specs — browse specification documents; opening one gives it its own tab, keyed by document id so it survives a rename.
Files — the project's workspace tree. Opening a file replaces the tree tab: the tree is a launcher, and in a narrow column a tab spent on the thing that only got you here is a tab wasted.
Loops — a project-wide glance at every loop. Opening one gives it a tab of its own, and the index stays open beside it.

Which tabs you have open is remembered per project; the panel's width is a single preference shared across all of them. Closing the last tab closes the panel and returns the full width to the conversation.

Jobs and loops¶

View and manage scheduled recurring work:

Jobs list — all scheduled jobs with cron expressions and next run times
Job history — recent execution results and status
Toggle controls — enable or disable jobs directly in the UI
Manual trigger — run a job immediately outside its schedule

A job with a stated purpose and a stop condition is a loop, and it gets more than a job does. Its drill-down tab shows the purpose, the stop condition, the queue by status, the item claimed right now, open questions, and the firing history — plus a live indicator while a firing is in progress.

Because every firing opens a new conversation, an active loop would otherwise fill an agent's conversation list with threads you never started. Two things prevent that:

A conversation created by a firing names the loop it came from, and clicking that name opens the loop's own tab. A plain scheduled job's conversation shows nothing — it has no loop.
Consecutive firings of one loop collapse into a single expandable row. A firing waiting for you still says so on the collapsed row, and the row opens itself if it holds the conversation you are reading.

Editing a running loop is safe: the edit is staged, not applied. The loop tab shows each staged value beside the one still in force, labelled "In force now" and "From the next firing", and a firing already running keeps the definition it was briefed with.

Neither a job nor a loop can be deleted — both are archived, so the record of what ran survives. A loop that is still running cannot be archived at all.

Agent Cards¶

Connected agents are auto-discovered from your session. Each card shows:

Agent name and status
The bound runner and the model it resolves to
How full its context is, as a compact bar
Its most recent status message, when it has one

Roles are gone — an agent's behaviour comes from the charter bound to it, and its execution capability from the runner bound to it. Permission posture is set per run from the composer, not as a property of the agent.

An agent can carry a default posture for runs that state none — a peer message or a scheduled job has no composer to choose one. See Permission postures for what each posture checks; a workspace is a working directory rather than a wall, and a write that leaves it is recorded rather than prevented.

Tips¶

Keep the dashboard open in a separate tab while agents are working
Use the tasks board to track progress without reading every message
Answer human questions promptly — agents may be waiting for your input before continuing
Check agent output logs when debugging issues
Use per-session chat to maintain separate conversation threads with each agent