But it then connects with and distributes work to both agents.
It also ignores the cpuMultiplier property set on the agent (set to 0.75), despite the Horde web interface showing LogicalCores as 27 (PhysicalCores is 36) there are 30 simultaneous tasks executing on that agent.
Is there another configuration setting needed to ensure UBA only distributes tasks to the specified pool and respects the number of cores?
The log line Pool: HelperOnly only shows that the provider parsed the ini. It does not mean Horde leased only that pool.
Two things are easy to mix up.
The machine you start the build on always runs tasks itself. Pool does not select it. If one of the two agents you see is the workstation running Visual Studio, that one is expected even though it sits in Win-UE5. The pool only filters remote helpers.
A custom property in agent.json is not membership in a Horde pool. Writing a property there does not put that agent in the HelperOnly pool, and it does not stop the other agent taking work. In the Horde UI, open each agent and check Pools. The helper should be in HelperOnly only, and the workstation agent should not be in that pool. The string in the ini has to match the pool id exactly. If HelperOnly has no eligible agent at the moment UBA asks for a lease, Horde can hand the work to any connected agent in the cluster, which looks exactly like the pool being ignored. The compute session on the server shows the pool it actually assigned. That is the line to compare with the ini.
cpuMultiplier is a separate knob. LogicalCores 27 (0.75 times 36) is what Horde’s own job scheduler reports. The UBA agent does not cap itself to that number. It will run more compile actions than cores, because those processes are not all busy at once, so 30 tasks on a 27-core agent is oversubscription rather than the multiplier being skipped. The MaxCores of 576 in your provider ini is the remote lease cap, not a per-agent cap. To hard-limit that box, cap workers or set processor affinity on that agent.