HarnessConfigLoader

Loads a org.llm4s.agent.harness.HarnessProfile from llm4s.agent.harness at the application's edge (design §4.37): the org.llm4s.agent.harness.HarnessProfile.deepAgents preset with the configured settings - which tools, the tool timeout, the prompt mode, each feature's limits or enabled = false, and the retention. llm4s-agent's reference.conf holds the defaults, which are the preset's, and the application's application.conf or -D system properties override them.

for
 profile <- HarnessConfigLoader.load()
 agent   <- Agent.builder("researcher", client).withHarness(profile.withPermissions(rules)).build()
yield agent

The loaded profile is checked as AgentBuilder.build() checks it, so a bad value is a ConfigurationError here - except that execute may be selected before the sandbox it needs is set. Backends, a sandbox, permission rules and a model compressor are code: set them on the loaded profile.