Refactor agent interface a bit

[rbren]

Right now, the LangChain agent also includes all its own logic for running commands, browsing, reading/writing files, etc. That needs to get factored out so that all agents can use it.

From what I can tell from the recent refactor, we basically just instantiate an agent with a task, and call run(), then it continues until it finishes.

This refactors things so that we have a flow more like this:

• server receives a task request
• server instantiates an agent
• server starts a control loop:
  • in each iteration, call agent.step()
  • agent.step() is allowed to send at most 1 LLM prompt
  • agent.step() returns an action, one of:
    • read file
    • write file
    • run cmd
    • browse
    • no-op
  • looper executes the action
    • and optionally passes messages on to e.g. the client browser

[neubig]

This looks great to me! I also looped in @xingyaoww in case he wants to take a look before we merge this in.