A new pattern is emerging.

The frontier models are becoming increasingly good at spawning and managing other long running agents.

An orchestrator sits at the top.

You discuss high level details with the orchestrator and it dispatches work on your behalf.

This saves context for the orchestrator.

The orchestrator also needs to wake itself up to check on tasks progressing over many hours/days. But you don’t want to waste its context on a null check. It also dilutes the conversation history for the human operator.

And you need a book keeper to keep notes organized so other agents can read from and possibly write to the project notes.

This goes beyond compaction of a single agent.

A lot of the core primitives come from the OpenClaw model.

But there is something else emerging here.

Grok Bot is a great example of these types of agents systems. It becomes especially obvious if you are using voice mode. The fast model constantly calls out to a slower model to do any heavy lifting. Somewhere behind the hood some system of agents handle task management, memory management, and preferences.

The result is not a dumb agent swarm, but an intelligent agent system.

Depending on how it’s designed you can dip into and out of the details as you see fit.

The problem of working with an orchestrator is that you are playing a game of telephone with the workers.

Some of your intent can get lost when you delegate the prompting to your orchestrator.