Breaking Down Tasks (Task Decomposition) › When Decomposition Goes Wrong

Task decomposition is powerful, but it's not foolproof. Let's look at some common issues and how to address them.

Too Vague

Sometimes the agent creates subtasks that are still too broad:

Bad decomposition:

1. Do research
2. Create content
3. Finish project

These aren't specific enough to guide action. You want concrete, actionable steps. To fix this, prompt for more detail or give examples of the level of specificity you want.

Wrong Order

The agent might list steps in an illogical sequence:

Bad decomposition:

1. Book flights to Paris
2. Decide where to go for vacation
3. Pack bags

Obviously, you need to decide where to go before booking flights. If this happens, you can either prompt the agent to reconsider the ordering or manually reorder the steps.

Missing Steps

The agent might overlook important subtasks:

Incomplete decomposition for "Host a dinner party":

1. Send invitations
2. Cook meal
3. Clean up after guests leave

What about planning the menu? Buying groceries? Setting the table? If you notice gaps, you can prompt the agent to expand its list or add missing steps manually.

The good news is that as models get better at reasoning, these issues become less frequent. And even when they occur, having some plan (even an imperfect one) is usually better than having no plan at all.