Workflow → Agent Evolution

Every agent pattern evolves from a workflow. When conditional logic becomes too complex to hardcode, the LLM takes over flow control.

Workflow

Prompt Chaining

Sequential LLM calls with validation gates between steps.

Workflow

Parallel Calls

Concurrent LLM calls on independent inputs, aggregated at the end.

Workflow

Orchestrator-Worker

LLM decomposes a task and delegates to specialized workers.

Workflow

Evaluator-Optimizer

Generate-evaluate feedback loop that iteratively improves output.

Agent

ReAct

Reason-act loop: the LLM reasons, calls a tool, observes, and repeats until done.

Agent

Plan & Execute

LLM creates a full plan upfront, then executes each step sequentially.

Agent

Tool Use

Structured function calling with schema-validated tool dispatch.

Agent

Memory

Persistent state across sessions: short-term, long-term, and semantic memory.

Agent

RAG

Retrieval-augmented generation: retrieve relevant context before generating.

Agent

Reflection

LLM critiques its own output and self-improves through structured feedback.

Agent

Routing

Intent classification dispatches inputs to specialized handlers.

Agent

Multi-Agent

Supervisor-worker delegation across multiple autonomous agents.

Workflow

Agent Pattern

Hover to highlight · Click for details

Workflow	Evolves Into	What Gets Added
Prompt Chaining	ReAct	Dynamic tool selection + LLM-controlled looping replaces hardcoded steps
	Tool Use	Structured function schemas + dispatcher replaces manual output parsing
	Memory	Persistent state store + retrieval replaces ephemeral context
Parallel Calls	RAG	Vector search retrieval replaces static parallel splits
Parallel Calls	Routing	LLM intent classification replaces hardcoded branch logic
Orchestrator-Worker	Plan & Execute	LLM plans full sequence upfront vs. dynamic decomposition
Orchestrator-Worker	Multi-Agent	Workers become autonomous agents with their own tools & memory
Evaluator-Optimizer	Reflection	Evaluator becomes the LLM itself via self-critique prompting