Task Decomposition
Splitting one big, vague job into smaller concrete pieces an agent can actually tackle.
Think of it like
Turning “plan the wedding” into a checklist — venue, invites, catering — because nobody can do “plan the wedding” in one move.
Example
“Migrate the database” becomes: back up the data, create the new schema, write the transfer script, verify counts, cut over.
How it actually works
Decomposition is the first move in planning: big goals are too abstract to act on, so the agent breaks them into subtasks small enough to have a clear next step. Good decomposition makes the rest easy and bad decomposition dooms it — miss a step and the whole chain is off. It is one of the things models do decently but not reliably, especially when the pieces have hidden dependencies.
For product teams
Whether an agent can handle a big ask usually comes down to how well it carves it up.
For engineers
Breaking a high-level goal into an ordered set of tractable subtasks, respecting dependencies.
Related
- Planning — The stage decomposition kicks off.
- Subtask — The pieces it produces.
- Supervisor Agent — The boss agent that often does the carving.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome