In depth
An agentic workflow replaces rigid scripts with autonomous LLM-driven execution. Instead of 'do A, then B, then C', the agent is given a goal and a toolbox, and it figures out the path. This adaptiveness is the defining feature — the workflow handles unexpected inputs, partial failures, and unknowns that break traditional automation.
Typical stages: **decompose** (break goal into sub-tasks), **delegate** (assign each sub-task to a tool or sub-agent), **execute** (call tools, observe results), **evaluate** (did we make progress?), **replan** (adjust strategy if stuck). Loop until the goal is met or a budget is exhausted.
Anthropic's published research (Building Effective Agents, 2024) outlines common patterns: **prompt chaining** (sequential steps), **routing** (classify then dispatch), **parallelization** (fan-out-fan-in), **orchestrator-workers** (manager agent + worker agents), **evaluator-optimizer** (generate + critique loop). Real systems combine these.
MCP is a natural fit for agentic workflows because MCP tools are the agent's action surface. Nanobot Enterprise, LangGraph, CrewAI, and AutoGen all wire MCP (or MCP-like) tools into their workflow engines.