Workflows
Understanding workflow patterns for multi-agent systems
Pick a workflow pattern and connect agents so they share context across steps.
from praisonaiagents import Agent, AgentFlow
researcher = Agent(name="Researcher", instructions="Research topics.")
writer = Agent(name="Writer", instructions="Write clear summaries.")
flow = AgentFlow(steps=[researcher, writer])
flow.start("Explain workflow patterns in one paragraph.")The user chooses a pattern, configures agents, and submits a task that flows through the team.
Quick Start
- Simple Usage
Chain two agents so the second builds on the first.
python from praisonaiagents import Agent, AgentFlow researcher = Agent(name="Researcher", instructions="Research topics.") writer = Agent(name="Writer", instructions="Write clear summaries.") flow = AgentFlow(steps=[researcher, writer]) flow.start("Explain workflow patterns in one paragraph.") - With Configuration
Run independent agents together with the top-level
parallelprimitive.python from praisonaiagents import Agent, AgentFlow, parallel ai = Agent(name="AI", instructions="Research AI trends") ml = Agent(name="ML", instructions="Research ML trends") summary = Agent(name="Summary", instructions="Combine research") flow = AgentFlow(steps=[parallel([ai, ml]), summary]) flow.start("Research latest developments")
How It Works
Workflows
Workflows define how multiple agents work together to complete a task. Just like a team of people in an office, agents can collaborate in different ways depending on the task.
What is a Workflow?
Think of a workflow as a recipe that tells your agents:
- Who does what
- In what order
- How they share information
The Four Main Patterns
“Send to the right expert”
Like a receptionist directing calls to the right department.
⚡ Parallel“Everyone works at once”
Like a team researching different topics simultaneously.
➡️ Sequential“One step at a time”
Like passing a document from person to person for review.
👔 Orchestrator“Manager delegates work”
Like a project manager assigning tasks to team members.
Which Pattern Should I Use?
| I want to… | Use this pattern |
|---|---|
| Send requests to different specialists | Routing |
| Do multiple independent things at once | Parallel |
| Process step-by-step (A → B → C) | Sequential |
| Let an AI decide how to break down work | Orchestrator |
Real-World Examples
Example 1: Customer Support Bot
A customer asks: “Why is my order delayed?”
| Pattern | How it works |
|---|---|
| Routing | Classify the question → Send to Shipping Agent |
| Sequential | Lookup Order → Check Status → Write Response |
Example 2: Research Report
You ask: “Create a report on AI trends”
| Pattern | How it works |
|---|---|
| Parallel | Research AI + Research ML + Research NLP (all at once) |
| Sequential | Research → Analyze → Write → Edit (step by step) |
| Orchestrator | Manager assigns: “You research, you analyze, you write” |
Two Ways to Build Workflows
PraisonAI offers two approaches:
| Approach | Class | Best for |
|---|---|---|
| Deterministic Pipelines | AgentFlow | You define exactly what happens |
| Dynamic Delegation | AgentTeam + hierarchical | AI manager decides |
Best Practices
Match the pattern to the task shape
Independent subtasks fit parallel; ordered steps fit sequential; expert routing fits routing. Choosing the shape first keeps flows simple.
Use AgentFlow for deterministic pipelines
AgentFlow(steps=[...]) runs exactly the steps you list. Reach for AgentTeam + hierarchical only when you want an AI manager to decide.
Import primitives from the top level
from praisonaiagents import parallel (and when, loop) uses the friendly re-exports, so samples stay short and portable.