In other words have them do one thing, well, and succinctly.
Take a look at this post from Lewis Metcalf at Amp. The way he’s working is that a thread only has a few user messages, and context stays well under the 200k limit and even the “dumb mode” threshold somewhere beneath that, say around 150k or so.
OK, but Unix tools are composable. This points up a really important feature I think any agent harness should have, which is something like Amp’s thread:handoff. We need to be able to summarize and ‘pipe’ a thread’s “output” to a new thread that will do something else with it.