All Docs
FeaturesSaaS FactoryUpdated March 20, 2026

Agent Workspace Console

Agent Workspace Console

The Agent Workspace is a real-time control centre for monitoring and intervening in AI agent pipelines across your entire product fleet. It is accessible from the main dashboard sidebar under Agent Workspace (Terminal icon) at /dashboard/agent-workspace.

Overview

The workspace gives you a live view of every pipeline currently running across all your products, the individual agent jobs within each pipeline, and a streaming log feed — all in one place. You can observe passively or take direct control of any pipeline at any time.

Layout

The console is split into two panels:

Live Agent Panel (Left)

Lists all active pipelines. For each pipeline you can see:

  • Pipeline status — current stage (e.g. implementing, testing, awaiting_approval)
  • Project name — which product this pipeline belongs to
  • Agent job pills — individual agent jobs within the pipeline and their statuses
  • Elapsed timer — how long the pipeline has been running
  • Pending intervention badge — count of unresolved interventions awaiting acknowledgement

Clicking a pipeline card filters the Activity Feed Console to show only that pipeline's logs.

Activity Feed Console (Right)

A terminal-style, scrolling log stream. By default it shows log entries from all active pipelines. Selecting a pipeline in the left panel filters the feed to that pipeline only.

Quick Inject Bar

Located below the activity feed. Type a free-form instruction and press ⌘↵ (or the equivalent keyboard shortcut on your platform) to inject it directly into the selected running agent's context.

Human-in-the-Loop Controls

Each pipeline card in the Live Agent Panel exposes five intervention actions:

ActionDescription
Interrupt (⏸)Requests the agent to pause before executing its next tool call. The pipeline enters a paused state and waits.
Inject (💬)Sends a free-form instruction into the agent's active context. The agent incorporates the instruction into its next reasoning step.
Take Over (✋)Marks you as taking manual control of the pipeline. Auto-resume is blocked until you explicitly resume.
Resume (▶)Un-pauses a pipeline that has been interrupted or taken over. The agent continues from where it left off.
Note (📝)Attaches an observer note to the pipeline for audit or record-keeping purposes. Does not affect agent behaviour.

All interventions are recorded in the agent_interventions table with their type, message, status, and timestamp.

Real-Time Data Stream

The workspace is powered by a Server-Sent Events (SSE) connection to GET /api/agent-workspace/stream.

Connection Behaviour

BehaviourDetail
Initial snapshotOn connect, the stream immediately sends a fleet_snapshot event containing the current state of all active pipelines and up to 300 recent log entries.
Poll intervalThe server polls for changes every 3 seconds.
Diff strategyFingerprint-based — only pipelines whose status, job statuses, or pending intervention count have changed are included in fleet_update events.
Log streamingNew log entries are detected and emitted as fleet_logs events.
HeartbeatA heartbeat event is sent every 15 seconds to prevent proxy or load balancer timeouts.
Max pipelines tracked50 most recently started active pipelines.

SSE Event Types

EventPayloadDescription
fleet_snapshotPipelineSnapshot[]Full initial state of all active pipelines
fleet_updatePipelineSnapshot[]Changed pipelines only (diff since last poll)
fleet_logsAgentLog[]New log entries from active pipelines
heartbeat{ ts: number }Keep-alive ping
error{ message: string }Stream-level error

Authentication

The SSE endpoint requires an active authenticated session. Unauthenticated requests receive 401 Unauthorized. The stream automatically scopes data to the projects accessible to the authenticated user.

Tracked Pipeline Statuses

The workspace tracks pipelines in any of the following statuses:

queued · researching · architecting · designing · implementing · testing · awaiting_approval · releasing · marketing · documenting · syncing_branches

Pipelines that have completed, failed, or been cancelled are not shown in the live panel.

Audit Log

Every intervention action is persisted to the agent_interventions database table. This provides a complete audit trail of all human-in-the-loop activity, including who acted, which pipeline was targeted, what type of intervention was applied, and what message (if any) was included.