Skip to main content
Agents track

Agent or Workflow?

Most things built as agents should have been scripts. Five questions to decide before you write any code.

0 of 5 answered
Can you draw the decision tree for this task?

1.Can you draw the decision tree for this task?

If you can write down every branch the task could take, you can encode those branches directly.

Roughly what does one run cost in tokens?

2.Roughly what does one run cost in tokens?

Agents pay for exploration. Below roughly 30,000–50,000 tokens, that overhead rarely earns its keep.

What happens when it gets something wrong?

3.What happens when it gets something wrong?

Autonomy multiplies both good and bad outcomes. The question is whether a mistake is recoverable.

What is actually slowing this task down?

4.What is actually slowing this task down?

If the bottleneck is not judgement, an agent will not remove it.

Can the task check its own work?

5.Can the task check its own work?

Tests, type checks, linters and builds give an agent a signal to iterate against.

The criteria here follow the evaluation checklist in Anthropic's “How We Build Effective Agents”. The scoring is ours, and is meant to surface the trade-offs rather than settle them.