Skip to main content

​
Summary

Workflows execute a predefined sequence. Agents pursue a goal under changing conditions. Most real systems sit somewhere between those extremes.

​
Why It Matters

Many product and engineering mistakes come from choosing the wrong control model.
  • If the task is stable, using an agent can add cost and unpredictability.
  • If the task is ambiguous and changing, using only a rigid workflow can make the system brittle and high-maintenance.
Choosing the wrong pattern is often more damaging than choosing the wrong model.

​
Mental Model

The cleanest distinction is about who owns the next step.
  • In a workflow, the designer owns the next step. The system follows a scripted path with explicit branching.
  • In an agent, the runtime owns the next step. The system chooses actions based on the current state, tools, and goal.
This does not mean workflows are simple and agents are advanced. It means they solve different coordination problems. Workflows are strongest when:
  • the path is known
  • the rules are stable
  • auditability matters more than flexibility
Agents are strongest when:
  • the path is not fully known in advance
  • the system needs to search, explore, or adapt
  • the environment can change while the task is running

​
Architecture Diagram

​
Tool Landscape

Most useful products are hybrids rather than pure examples of either side.
  • A workflow may call an agent for one ambiguous step.
  • An agent may operate inside a larger workflow with strict entry, approval, and exit points.
  • A research or coding system may use workflow-like stages but agentic decision-making inside each stage.
That hybrid design is often the practical answer because it preserves deterministic structure where structure helps, while reserving autonomy for the parts that genuinely benefit from it.

​
Tradeoffs

  • Workflows are easier to test and audit, but they are expensive to maintain when edge cases keep growing.
  • Agents adapt better, but they need stronger safeguards, observability, and fallback design.
  • Hybrid systems are often best in practice, but they require clearer interface boundaries than either extreme.
Useful defaults:
  • choose workflows for deterministic business policy
  • choose agents for bounded exploration and judgment
  • wrap agents with workflow controls when compliance, approvals, or irreversible actions matter

​
Citations

​
Reading Extensions

​
Update Log

  • 2026-04-21: Initial repo-native draft based on imported reference material and lab rewrite rules.