We ran into something that didn't seem like a problem until it was. Each agent had access to the tools it needed and everything worked fine in isolation.
The issues started once agents were running in parallel. Two parts of the system would try to use the same tool or hit the same resource at the same time. Results became inconsistent and it wasn't obvious why.
Limiting access helped in some cases but slowed things down elsewhere. Too much access caused race conditions. Too little caused steps to stall waiting for something to free up.
Most of the coordination logic ended up sitting outside the agents themselves. Every new agent added more decisions around what it should be allowed to access and when.
There isn't a shared way to manage tool access across a multi agent system. How are you handling this when multiple agents are running at the same time?
[link] [comments]