Context should follow the work

We work through coding tasks one step at a time, but a software engineer’s mental model usually looks more like a call stack: push a task on, work through it, then pop it off when it is complete.

In December 2025, I thought the standard linear chat history was the wrong shape for this. A long-running agent could move from task to task while carrying the conversation forward as one growing sequence.

I wanted its context to follow the structure of the work. A call stack, or a flame graph of that stack, seemed a more useful model.

My hypothesis was that organizing context this way could avoid compaction in some cases and make it less lossy when it was necessary. A larger context window alone would not address the architectural question.

That also suggested a role for an intelligent agent to orchestrate the subagents and manage a changing call stack.