How Andrew Ng's OpenWorker Orchestrates Context
Reading a large codebase fills an agent's context. OpenWorker hands the job to a explorer with a fresh window, and keeps only the report it comes back with.
The interesting part is not the usual model → tool → result loop. Most agent frameworks already do that.
The part where I think you should focus at is how OpenWorker handles context.
When the main agent needs to inspect a large codebase, it can call an explore tool. That creates another TurnEngine with a fresh context window.
Now you should be thinking, what is thisss explore tool, and what exactly is a TurnEngine? 🤯
So, in a very simplistic sense, TurnEngine is the thing running the agent loop. It takes the current conversation, sends it to the model, executes whatever tool calls the model asks for, feeds those results back, and keeps doing this until the task is complete or the iteration limit is reached.
And explore is simply one of the tools available to that main agent. The only difference is that instead of reading a file or running a search directly, calling explore spins up another TurnEngine.
Basically:
The explorer gets read-only tools for things like file search, file reading and git inspection. It can go through a large amount of code without polluting the main agent’s context. Once it is done, only the final report comes back to the parent agent.
You can loosely relate this to exploration vs exploitation.
The main agent keeps working with the information it already has. You can think of that as exploitation.
When it needs more information, it spins up an isolated explorer, spends a separate context window there, and brings the useful parts back. That is the exploration side.
This is not exploration/exploitation in the strict RL sense. There is no learned value function or reward-based policy involved.
The comparison is more about how OpenWorker allocates context.
OpenWorker is deciding whether to keep working with the context it already has, or spend a fresh context window to go find more information.
And that, I think, is the more interesting way to look at its context orchestration.
Sources
Original Github Repo : https://github.com/andrewyng/openworker
