Add configurable limit for parallel agent execution (or API rate limits)

[DetouristCollective]

Directing this issue on the cline repo to the appropriate channel (kanban repo):

When adding many tasks to the backlog and clicking play, all agents start running simultaneously. This can overwhelm system resources and API rate limits.

It would be useful to have a configurable setting that controls the maximum number of agents that can run in parallel. For example, a user could set a concurrency limit of 5, so only 5 agents execute at a time, with the remaining tasks queued until a slot opens up.

Proposed behavior:

    Add a setting (e.g. cline.maxParallelAgents) to control the maximum number of concurrent agents
    Default to a reasonable limit (e.g. 3-5) rather than unlimited
    Tasks beyond the limit are queued and start automatically as running agents complete

In addition to @leighmcculloch's suggestion, I'd also like to add:

Since API rate limits are the primary bottleneck, instead of trying to mitigate it by proxy with maxParallelAgents, adding an API rate limit setting (e.g. set n and t so that agents wait if it hits n requests in a rolling t seconds window) may result in the maximum task throughput while minimizing API request failures.

However, I suspect that this is blocked by cline/cline#588 (cline agent's ability to pace itself based on API rate limits)

[leighmcculloch]

Since API rate limits are the primary bottleneck

I hadn't considered this motivation, that does seem important too.

The primary reason I opened the original issue was diminishing returns. If more than a reasonable number of issues get worked on there is increasing chance I end up with conflicting work, rebase loops, and the review list gets too long and too far away than when I kicked off the issue and had context. There might be other things I can do about my workflow to address those issues, but so far what I've found most effective at finding that sweet spot is to simply limit my parallelisation to something where those downsides are minimal.

[Bennett-Wendorf]

As a part of this, I wonder if it might be useful to have a swimlane for "Queued", so that I can queue up a number of tasks to run, but limit how many are running at once.

I'm testing out Cline Kanban against a local agent, and as such have very limited compute resources. I can't easily run more than a couple tasks simultaneously, but it's often viable in my workflow to have 8-10 tasks ready to go. I would love if I could queue up all 8 and then have them execute 1 or 2 at a time, with the next one only starting when there is a slot available in "In Progress".