Delegation
The next time you ask your computer to do something, try saying: “use sub-agents.”
When an AI model uses sub-agents, it can either do the work itself or delegate parts of the work to other AI threads. The point is not to add ceremony. The point is to match the difficulty of each task to the level of reasoning it needs.
Delegate the easier work
For a large project, use a high-reasoning model as the orchestrator. Give it the goal and the full context, then ask it to delegate work that is clear and less complicated to agents at a lower reasoning level.
This can save tokens while producing consistent results. You give up some control, but you gain a practical way to reserve expensive reasoning for the parts of the project that need it.
Have sub-agents debate
When you are not sure what the core problem is, ask an orchestrator to have sub-agents debate it. “Adversarial review” is one name for this, but “debate” is enough.
The sub-agents can propose different explanations, challenge one another, and surface assumptions. The orchestrator then consolidates the discussion into a likely conclusion and a decision for you to review.
Your job is judgment
Delegation does not remove responsibility. Read the conclusion, decide whether the reasoning is sound, and choose what happens next. You do not need to read every line of every debate, but you do need to judge the result.
Use delegation to improve plans
Sub-agents are also useful for planning. Ask several agents to create a plan, improve an existing plan, or find a simpler approach. Then let the orchestrator combine the strongest ideas and reserve the final decision for you.
Orchestration is a useful mental model: delegate the work that is easy to separate, and keep the hardest judgment close to the person responsible for the outcome.