[refactor] better agent abstraction

It would be good to abstract the 3 different (client side) agents (concurrency, activation db and ngrok) using an `Agent` class, and fully encapsulating the logic inside Agent. For example, that ngrok is working very different from the others and does not use a polling loop. That loop is currently in `Debugger.run()`, but it should not have to know about these details.

Requirements:
- `Agent` base class
- 3 different implementations in `src/agents/`
- move logic how to install agent specific things from `AgentMgr` into agent (exception: if installing concurrency fails, install activation db agent #7)
- run loop in agent, gets callback for invoker.run() passed in

Start by reading AgentMgr and Debugger.run() to trace installation and the current polling loop, then inspect the three client-side agent paths under src/agents/. The work is done when a base Agent and three implementations own agent-specific installation and run behavior, while the concurrency fallback and invoker.run() callback remain functional.

Written by the indexing model from the issue text.