Study · Claude Course

Agents and Workflows

A **workflow** is a predetermined sequence of calls (use when steps are known); an **agent** plans dynamically with a toolset (use when steps are unclear). Prefer workflows for reliability; reach for agents only when flexibility is truly needed.

type conceptstatus activeagents · workflows · patterns · architecture

Key points

  • Workflow patterns: evaluator-optimizer (producer generates, evaluator scores, loop); parallelization (split into simultaneous subtasks, then aggregate); chaining (sequential single-focus steps); routing (classify input, then route to a specialized pipeline).
  • Agents: plan dynamically; prefer abstract, generic tools (bash, web_fetch, file_write) over hyper-specialized ones; a small flexible toolset works best.
  • Environment inspection: agents check results after each action (screenshot after a click, read a file before editing) to adapt instead of acting blindly.
  • Trade-off: workflows = testable, higher success, need specific inputs; agents = flexible, harder to test, lower success. Reliability first, innovation second.
Tool-Use · Model-Context-Protocol · Retrieval-Augmented-Generation

Sources

  • 2026-06-28-claude-course
Compiled from wiki/study/claude-course/Agents-and-Workflows.md · git is the source of truth