AI Engineering · interactive

Agent patterns: workflow or agent?

Most “agents” in production are workflows: model calls wired together by your code. A true agent chooses its own next step in a loop. Run six patterns from Anthropic's Building Effective Agents and compare the cost of each.

Input

product notes

LLM call 1

draft

Code check

< 120 words?

LLM call 2

translate → es

Prompt chaining: The task splits into fixed steps, and each step is easier than the whole.

Pattern

model calls: 2
wall time: ~4s
predictable: high

Task: Write a launch email, then translate it to Spanish.

Step 1 of 3 · read left to right

1
Draft the email
One call does one job: write the English draft.
Draft: 96 words.
2
Check it in code
A gate between calls catches a bad draft before it costs a second call.
96 < 120. Pass.
3
Translate it
The second call only sees the checked draft.
Spanish email ready.

Compare all six

Start at the top. Move down only when the simpler pattern measurably fails.

Who decides the path	Calls	Wall time	Predictable
Your code	2	~4s	high
Your code	2	~4s	high
Your code	3	~2s	high
A model plans, code runs it	5	~6s	medium
Your code	4	~8s	medium
The model, every turn	4	~8s	low

In production

Tools and technologies

Agent frameworks are mostly ways to wire these patterns; the patterns work with a plain model API too.

Not tied to one language — use these names when you search docs or design a system.