Workflows

Purpose: Compose multi‑agent pipelines. Helper agents run first (in parallel), then their outputs are interpolated into a final prompt for the primary agent.

Concepts

• Main Agent: the final agent that produces the delivered output.
• Helper Agents: upstream agents that generate structured inputs (facts, summaries, lists) for the main agent.
• Data interpolation: the workflow's mainPrompt can interpolate both helper outputs and your supplied data.
• Parallelism: helper agents run in batches (set by concurrency).

Interpolating data, not just agents

Workflows now interpolate plain data alongside agent outputs. You supply a JSON object as the workflow's dataPayload (and can override it per run), then reference it in mainPrompt two ways:

• Data templates {{ path }} — replaced inline with a value from dataPayload (text, or pretty‑printed JSON for objects/arrays). Supports dotted paths and array indices, e.g. {{ audience.segment }} or {{ sources.0.title }}.
• Helper selectors <selector|#|dataKey> — run a helper agent and substitute its output. The dataKey can itself be a {{...}} template, a top‑level data key, a comma‑separated list, or a { "scrape_url": "https://..." } object (the page is fetched and inlined).

The selector delimiter defaults to |#| and is configurable per workflow.

Design tips

• Keep helper agents single‑purpose (e.g., “gather latest headlines”, “extract tickers”).
• Use Prompt Studio to harden each helper before composing.
• Prefer explicit placeholders in templates; avoid implicit concatenation.
• Validate helper outputs (length, format) before interpolation to reduce prompt bloat.

API parity

• Saved configs + resumable runs: POST /v1/core/workflows, POST /v1/core/workflow-runs
• Legacy immediate execution: POST /v1/core/agent-workflow

See Workflows (API). To chain workflows and agents into longer, schedulable pipelines, see Factories.