Improvements to providers concept

[tomkis]

Let's use this ticket as a

Current limitations of Providers concept

• Env variables cannot be scoped to agents, there is now way how I can have 1 provider with 2 agents, each using different models)
• Provider is tightly coupled with agent - agent's dependencies and installed and run with provider, providers are not lightweight as intended
• Agent installation is happening magically under the hood - it should happen at users request (agent installation might fail, essentially fails all other agents within the provider; like with ollama, user should be able to see the installation and it's progress)
• Model and other dependencies should be installed only when user needs the agent - for example if agent depends on llama 3.3 from ollama, I only want this dependencies to be installed on user request
• deployments - we need to be able to deploy agents as standalone docker containers, not a group of agents within provider; also, I think the manifest should follow k8s manifests (envs, configmaps, secrets)

Resource Utilization

Runtime Configuration

Security