You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Keep one long-lived agent per client on a single workstation. I do freelance infrastructure and ops work, so at any time there are several unrelated clients in flight, each with its own directory and its own ongoing conversation. A session stays useful for weeks. Some get touched every day; others sit untouched for a week and then I pick them up exactly where they were, because the accumulated context is the whole value.
How I do it today
I leave them all open. Every resident agent holds a provider process, so the number of clients I can have "live" at once is set by how much RAM I have, not by how much work I actually have. When I run out of headroom I close whichever ones I guess I'm least likely to need next, which is a guess I'm often wrong about.
Where Paseo gets in the way
Closing and archiving are both decisions about whether work is finished. What I actually want to express is "not right now, probably Thursday". Archiving takes the agent out of the active list, so the only lever I have for reclaiming memory is also the one that removes the thing from where I work. The result is that memory pressure quietly pushes me toward pretending work is done when it isn't.
To be clear about what I'm not asking for: the automatic idle collector that was removed in #2590 was, I think, correctly removed — it collected runtimes at moments the user didn't choose and couldn't see, and that lost messages (#2568, which I hit myself on 0.2.3). A timer is the wrong shape for this. The thing I miss is a deliberate action.
What would the flow look like if it worked for me
I'd right-click an agent I know I won't touch today and park it. It stays in the sidebar where I left it, visibly parked rather than idle, and its process is gone. When I come back — next prompt, or just opening it — it wakes through the provider's normal resume and carries on. Waking would be slower than an already-resident agent, and that's exactly the trade I want: a little latency on next use in exchange for the memory back now.
The bit that matters most is that parking is not a filing decision. The agent doesn't move, doesn't leave the list, and doesn't need to be found again afterwards.
I originally wrote this up as an issue (#2619) before reading CONTRIBUTING properly — reposting here in the right shape.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
What am I trying to do
Keep one long-lived agent per client on a single workstation. I do freelance infrastructure and ops work, so at any time there are several unrelated clients in flight, each with its own directory and its own ongoing conversation. A session stays useful for weeks. Some get touched every day; others sit untouched for a week and then I pick them up exactly where they were, because the accumulated context is the whole value.
How I do it today
I leave them all open. Every resident agent holds a provider process, so the number of clients I can have "live" at once is set by how much RAM I have, not by how much work I actually have. When I run out of headroom I close whichever ones I guess I'm least likely to need next, which is a guess I'm often wrong about.
Where Paseo gets in the way
Closing and archiving are both decisions about whether work is finished. What I actually want to express is "not right now, probably Thursday". Archiving takes the agent out of the active list, so the only lever I have for reclaiming memory is also the one that removes the thing from where I work. The result is that memory pressure quietly pushes me toward pretending work is done when it isn't.
To be clear about what I'm not asking for: the automatic idle collector that was removed in #2590 was, I think, correctly removed — it collected runtimes at moments the user didn't choose and couldn't see, and that lost messages (#2568, which I hit myself on 0.2.3). A timer is the wrong shape for this. The thing I miss is a deliberate action.
What would the flow look like if it worked for me
I'd right-click an agent I know I won't touch today and park it. It stays in the sidebar where I left it, visibly parked rather than idle, and its process is gone. When I come back — next prompt, or just opening it — it wakes through the provider's normal resume and carries on. Waking would be slower than an already-resident agent, and that's exactly the trade I want: a little latency on next use in exchange for the memory back now.
The bit that matters most is that parking is not a filing decision. The agent doesn't move, doesn't leave the list, and doesn't need to be found again afterwards.
I originally wrote this up as an issue (#2619) before reading CONTRIBUTING properly — reposting here in the right shape.
All reactions