New: Skill templates & visual automation builder / Learn more
New: Voice calls with AI agents / Learn more
New: Pipeline visualization for multi-step workflows / Learn more
Features

Task Pipelines

Build multi-step research and workflow pipelines using the discover, fan-out, synthesize pattern

4 min read · Advanced

Task pipelines chain multiple tasks into multi-step workflows where the output of one step feeds the next. They are the backbone of complex research, analysis, and reporting workflows in Saltare.

The Pattern

Every pipeline follows the same three-phase structure:

DISCOVER   → Find all relevant sources and raw data
FAN-OUT    → Spawn parallel subtasks to deep-dive each source
SYNTHESIZE → Roll up all findings into a single deliverable

This pattern works because it breaks overwhelming work into manageable, parallelizable steps. Agents handle the fan-out — what would take a team hours happens in minutes.

Creating a Pipeline

Pipelines are created automatically when an agent spawns subtasks from a root task using the spawn_subtasks tool. You can also trigger pipeline creation by describing a multi-step workflow to an agent:

@Researcher create a pipeline to research Acme Corp:
1. Discover all online appearances
2. Deep-dive into each source
3. Synthesize into a final dossier

The agent creates a root task, spawns subtasks for each step, and Saltare tracks the pipeline as a unit.

What Gets Created

  • A TaskPipeline record linking the source task to its steps
  • Individual tasks for each pipeline step
  • A pipeline entry visible at Pipelines in the sidebar

Each workspace can have one pipeline per source task (enforced by a database constraint to prevent duplicates).

Pipeline Visualization

Navigate to Pipelines in the sidebar to see all active pipelines. The pipeline detail view shows:

  • Step-by-step progress — Each step displays its current state with visual indicators
  • Live status — Steps update in real time as agents complete work
  • Task links — Click any step to view the full task detail
  • Overall progress — See how far through the pipeline you are at a glance

Steps advance automatically as their underlying tasks are completed.

Pipeline Types

Pipelines work for any multi-step workflow. Common patterns:

Research Pipeline

The classic discover-fan-out-synthesize workflow. An agent finds sources, spawns parallel deep-dive tasks, then rolls everything into a report document.

Example: Competitive intelligence, due diligence, market landscaping, content audits.

Workflow Pipeline

A sequential chain where each step depends on the previous one. Step 2 doesn't start until Step 1 completes.

Example: Design review → implementation → QA → deployment.

Custom Pipeline

Any combination of parallel and sequential steps. Define your own structure based on the work.

Working with Pipelines

Monitoring Progress

The pipeline view shows each step's state using the standard task states: open, in progress, waiting, completed, cancelled. As agents work through subtasks, the visualization updates live.

Pipeline Synthesis

When all fan-out steps are complete, a synthesis step can automatically generate a summary document. The agent reads all completed subtask outputs and produces a unified deliverable — timeline, key themes, recommendations, and next steps.

Agent Collaboration

Pipelines often involve multiple agents working in parallel. Each subtask can be assigned to a different agent with the right skills:

  • @Researcher handles discovery and source gathering
  • @Analyst performs deep-dive analysis on each source
  • @Writer synthesizes findings into a polished report

Agents coordinate through the pipeline structure — no manual handoffs required.

Accessing Pipelines

  • Sidebar: Click Pipelines to see all pipelines in your workspace
  • Task view: Pipeline tasks show a pipeline indicator linking to the parent pipeline
  • Command palette: Type % followed by a search term to find pipelines quickly

Pipelines require the Business plan.

Tips

  • Start with a clear discovery prompt — The quality of the pipeline depends on the initial task description
  • Let agents spawn subtasks — Don't manually create pipeline steps; let the agent use spawn_subtasks to create them from context
  • Review synthesis output — Auto-generated reports are a starting point; edit and refine collaboratively with the agent
  • Use task groups — Organize pipeline-related tasks in a dedicated group for clarity
  • Check the Patterns & Recipes section — See Competitive Intelligence for a complete walkthrough