Workflow Agents (ADK)

Workflow Agents are specialized, deterministic controllers in the ADK that manage the execution flow of other agents. Unlike LlmAgents, which use LLMs to decide their next action, Workflow Agents follow predefined patterns for orchestration.

Core Opinion

Workflow agents matter because they move flow control out of the model and into explicit structure. In the Nest, they are the clearest example of when "agentic" does not have to mean "LLM decides everything."

They are the right abstraction when:

• the execution pattern is already known
• predictability matters more than conversational flexibility
• you want model intelligence inside steps, not in the control plane itself

Core Characteristics

• Deterministic: Execution paths are fixed and predictable.
• No LLM Logic: They do not use language models to determine flow (though their sub-agents might).
• Orchestrators: Their primary role is to manage one or more sub_agents.

Types of Workflow Agents

1. Sequential Agent (SequentialAgent)

Executes a list of sub_agents in a strict, linear order.

• Use Case: Multi-step pipelines where Step B depends on the completion/output of Step A (e.g., Generate -> Review -> Publish).

2. Parallel Agent (ParallelAgent)

Executes multiple sub_agents concurrently.

• Use Case: Independent tasks that can be performed simultaneously to reduce latency (e.g., searching three different data sources at once).
• Note: Events from sub-agents may be interleaved in the session history.

3. Loop Agent (LoopAgent)

Repeatedly executes its sub_agents until a specific exit condition is met.

• Use Case: Iterative processes such as self-correction, data polling, or recursive decomposition.

Comparison: Workflow Agents vs. LLM Agents

Decision Rule

Start from workflow-agents when your question sounds like one of these:

• "Should this coordination pattern be explicit instead of model-selected?"
• "Which deterministic controller fits this multi-step job?"
• "How does ADK express sequence, parallelism, or iteration structurally?"

If the question is broader than ADK, route to graph-orchestration. If the question is about how agents share state or delegate, route to adk-multi-agent-orchestration.

Related Patterns

• multi-agent-systems: Workflow agents often sit at the top of a tree, managing specialized sub-agents.
• pattern-dynamic-delegation: While workflow agents use structural delegation, LlmAgents use dynamic delegation via tools or transfer.
• graph-orchestration: The higher-level pattern that workflow agents instantiate concretely inside ADK.

Start Here

1. Read this note if the main issue is choosing sequence vs. parallel vs. loop control.
2. Then read adk-multi-agent-orchestration if state sharing or transfer semantics matter.
3. Return to agent-development-kit if you need the broader runtime and service model.

*Source: lit-adk-documentation*