Chain LLM calls, agent tasks, data fetches, browser actions, and platform publishes into reusable pipelines. Run a product launch campaign, a weekly content series, or a custom research-to-publish workflow — triggered on demand or on schedule.
Brandstaq's skill execution engine runs on a 10-action instruction set. Every pipeline is a flow — a sequence of steps, each using one of these actions:
`llm_call` — Call an LLM with a prompt, get structured output. The core of most content generation steps.
`data_fetch` — Pull data from an API, database, or web source. Use this to feed real-time information (market data, analytics, competitor content) into your pipeline.
`agent` — Dispatch a specialist agent (Script Writer, Voice Agent, Image Agent) and get its structured output as input for the next step.
`platform_api` — Call a social platform API directly. Post to Twitter, publish to LinkedIn, upload to Instagram.
`browser` — Launch a headless browser to capture screenshots, scrape structured data, or interact with web pages.
`publish` — Publish a content job output to one or more platforms with full post type handling (text, image, video, thread).
`director` — Run a multi-agent production pipeline (like a full video generation workflow) as a single step.
`condition` — Branch the flow based on output content, metrics, or external data.
`loop` — Iterate over a list and run a sub-flow for each item.
`transform` — Reshape, filter, or reformat data between steps.
These actions compose into workflows of any complexity. Simple pipelines are 3-4 steps. Complex production workflows can have 20 steps with branching and loops.
Pipelines are not rigid scripts — data flows between steps dynamically through Brandstaq's variable resolution system. Every step's output is accessible to subsequent steps via `$steps.{stepName}.output`. Input parameters are accessible via `$input.{paramName}`. String templates use `{{input.topic}}` syntax for interpolation.
This means a pipeline can: fetch trending topics from Twitter ($steps.fetch.output.topics), pass the top topic into an LLM call to generate an angle ($steps.angle.output.recommendation), then pass that angle into a content writing step that produces a full blog post, then pass the blog post into a publish step that formats it for three platforms simultaneously.
Pipelines support parallel execution — steps with no dependencies on each other can run simultaneously, cutting total execution time significantly for workflows with multiple independent sub-tasks.
Pipelines in Brandstaq are stored as skills — reusable, versioned workflows that agents can discover and invoke during their daily work. Brandstaq ships with a library of seed skills covering common brand operations: social post generation, blog writing, voice preview, content repurposing, competitor monitoring, and more.
You can create custom skills through the admin interface by writing the flow JSON directly, or by asking Brandstaq to generate a flow from a description. AI-generated skills are validated for safety (cost cap of 50 tokens at generation time, `gen_` prefix enforced, deep validation of all handler references before execution).
Skills are versioned — when you update a flow, the previous version is preserved. Skills also have their own cost tracking via the `cost_tokens` field, so you know exactly what each pipeline invocation costs before running it. Prompts within skills optimize automatically via Brandstaq's DSPy bridge — MIPROv2-based optimization that improves prompt quality over time using your actual outputs as training signal.
No. You can use Brandstaq's admin interface to build pipelines visually, or describe the workflow you want in the chat interface and Brandstaq generates the flow. For complex pipelines with branching logic and custom data transforms, familiarity with JSON is helpful but not required. The seed skill library covers most common use cases out of the box.
Brandstaq enforces safety limits: maximum recursion depth of 5 (pipelines calling pipelines), maximum 20 total steps per run, maximum 50 steps defined in a single flow. Each step has a per-step timeout based on remaining worker time. These limits exist to prevent runaway executions and ensure every pipeline completes within the 10-minute agent window. For genuinely complex multi-stage workflows, use Brandstaq's Celery ETA dispatch to break them across multiple agent sessions.
Yes. The `data_fetch` action supports external API calls with authentication headers you configure in your brand kit. The `browser` action can access password-protected pages using session cookies imported from your browser. This enables pipelines that fetch from your internal analytics dashboards, CRM exports, or proprietary data sources before generating content.