Skip to content
Agentic WorkFlow

Advanced AI Workflows

Advanced AI workflows are normal workflows with one extra responsibility: they must keep model behavior observable and constrained. Use this section to decide when to add an LLM, how to pass it clean context, how to ground answers in retrieved data, and how to keep agent behavior testable.

AI Concepts

Learn the design ideas behind agents, embeddings, RAG, memory, model dependencies, prompting, and evaluation.

Explore Concepts →

LangChain Integration

Understand how LangChain-style components fit into Agentic WorkFlow nodes and dependencies.

Learn LangChain →

Browser AI Automation

Combine page extraction, browser actions, and model calls without losing track of page state.

See Browser Integration →

Multi-Agent Systems

Use agent patterns only when the workflow needs planning, tool choice, or iterative decisions.

Advanced Patterns →

When AI Belongs In A Workflow

Section titled “When AI Belongs In A Workflow”

Use AI when the workflow needs judgment over language or messy content. Do not use AI for work that a deterministic node can do more clearly, such as formatting a date, filtering a known field, or checking whether a number is greater than a threshold.

flowchart TD
  Need["Workflow need"] --> Known{"Can rules solve it?"}
  Known -->|Yes| Deterministic["Use flow, mapping, code, or transformation nodes"]
  Known -->|No| Language{"Is the input language-heavy or ambiguous?"}
  Language -->|Yes| AI["Use an AI node with constrained context"]
  Language -->|No| Redesign["Clarify data shape before adding AI"]
  AI --> Check["Validate output before downstream actions"]

Good AI steps usually have:

  • A clear task, such as summarize, classify, extract, answer, rank, or decide.
  • Clean input fields from previous nodes.
  • A model dependency that matches the task.
  • A structured output or parser when later nodes depend on the result.
  • A fallback path for empty context, model errors, or low-confidence answers.

Learning Path

Section titled “Learning Path”
  1. Start with Model Dependencies so you understand chat models, embeddings, vector stores, memory, parsers, and tools.
  2. Read Prompting and Structured Outputs before wiring model output into later automation.
  3. Use Embeddings and Vectors when you need semantic search.
  4. Use RAG when answers must be grounded in retrieved documents or page content.
  5. Use AI Agents and Tool Selection when the model must choose actions.
  6. Add Evaluation and Testing before relying on the workflow for repeated production tasks.

Common Patterns

Section titled “Common Patterns”
PatternUse it forMain risk
Extract then summarizeTurning page text into a shorter resultSending noisy or incomplete page content
Classify then branchRouting records based on language meaningVague categories that overlap
Retrieve then answerAnswering from a knowledge base or page collectionUnsupported answers when retrieval is weak
Agent with toolsMulti-step tasks where the next action depends on resultsToo many tools or unclear stop conditions
Structured outputPassing model results into later nodesMissing schema validation or fallback behavior

Design Checklist

Section titled “Design Checklist”
  • Keep the browser extraction step separate from the AI step.
  • Trim page content before sending it to a model.
  • Prefer structured output for anything used by another node.
  • Ground factual answers with retrieved context or source text.
  • Add an error path for empty input, credential failure, rate limits, and invalid output.
  • Test with short, long, irrelevant, and malformed inputs.

Where To Go Next

Section titled “Where To Go Next”