Agent Workspace Console
Agent Workspace Console
The Agent Workspace is a persistent, fleet-wide console that lets you watch every AI agent working across all your products in real time and intervene at any point without interrupting the overall pipeline.
Navigate to it via Sidebar → Agent Workspace (Terminal icon), or go directly to /dashboard/agent-workspace.
Layout
The workspace is split into two panels:
Left — Live Agent Panel
Shows every active pipeline run across all products. Each card displays:
- Product name and pipeline trigger
- Current pipeline status
- Per-agent job pills with individual statuses
- Elapsed time since the run started
- A badge for any pending human interventions awaiting action
Right — Activity Feed Console
A terminal-style scrolling log stream. By default it shows logs from all active pipelines merged together. Click a pipeline card on the left to filter the console to that run only.
Quick Inject Bar
Located at the bottom of the Activity Feed Console. Type any free-form instruction and press ⌘↵ (or Ctrl↵ on Windows/Linux) to inject it into the currently selected pipeline's agent context.
Human-in-the-Loop Controls
Each active pipeline card exposes five intervention actions:
| Action | Description |
|---|---|
| Interrupt | Requests the agent to pause before its next tool call. The pipeline enters a suspended state until resumed. |
| Inject | Sends a free-form text instruction into the agent's context at the next available step. The agent continues running but considers your instruction. |
| Take Over | Marks the pipeline as under manual control. Auto-resume is blocked — the pipeline will not continue until you explicitly resume it. |
| Resume | Clears an interrupted or taken-over state and allows the agent to continue from where it stopped. |
| Note | Attaches an observer note to the pipeline's audit log without affecting agent behaviour. |
All interventions are recorded in the agent_interventions table with their type, message, status, and timestamp, forming a persistent audit trail.
Real-Time Stream
The console is powered by a Server-Sent Events (SSE) connection to /api/agent-workspace/stream.
Connection behaviour
- On connect, the server sends an initial
fleet_snapshotevent containing up to 50 active pipeline runs and the 300 most-recent log lines. - Every 3 seconds, the server re-polls active pipelines. It uses fingerprint-based diffing and only emits a
fleet_updateevent when something has actually changed (status, job statuses, or pending intervention count). - New log lines are streamed as
fleet_logsevents as they appear. - A
heartbeatevent is sent every 15 seconds to keep the connection alive through proxies.
Event reference
| Event | Payload | Description |
|---|---|---|
fleet_snapshot | PipelineSnapshot[] | Full initial state of all active runs |
fleet_update | PipelineSnapshot[] | Partial update — only changed pipelines |
fleet_logs | AgentLog[] | New log entries from active runs |
heartbeat | { ts: number } | Keep-alive ping |
error | { message: string } | Stream-level error |
Authentication
The SSE endpoint requires an authenticated session. Unauthenticated requests receive 401 Unauthorized. The stream automatically scopes results to projects accessible to the authenticated user.
tRPC Router: agentWorkspace
The agentWorkspace tRPC router exposes 7 procedures for programmatic fleet management and intervention. These back the UI controls described above and can also be called from custom integrations.
Related
- Agent Marketplace — share, install, and rate agent packages
- Fleet Benchmarks — anonymized cross-product performance intelligence
- Audit Log — full intervention history