[Agent, Eval] Fixes LLM config issue for delegation & Add eval to measure the delegation accuracy

[xingyaoww]

What is the problem that this fixes or functionality that this introduces? Does it fix any open issues?

1. AgentDelegationAction does not re-use the LLM instance from the caller Agent, which makes it hard to accumulate cost metrics. It won't work correctly use the right LLM config during evaluation.
2. In the agent controller, if self.delegate is checked before the max iteration check, which makes the whole agent loop never stop.

---

Give a summary of what the PR does, explaining any non-trivial design decisions

1. Add LLM as an additional argument for AgentDelegationAction so that the child agent can re-use the same instance from the parent (that particular instance won't be saved to the event stream, though).
2. Fix the order of if self.delegate check and the max iteration check
3. add a benchmark (more like a test) to check whether the agent can successfully delegate the browsing task to the BrowsingAgent - this is needed to report CodeActAgent's browsing performance on the paper. README.

---

Other references

[neubig]

Basically LGTM, but just had one question to make sure that you intentionally removed the thought from the delegation task.

[neubig]

LGTM then!

[li-boxuan]

Moving this code around won't solve the real problem. The real problem is we don't accumulate steps across parents and children, and needs to be fixed by #2296.

This code move would likely introduce a bug: you will see the parent agent die while its child still running.

[li-boxuan]

Considering the slow progress of #2296, I can take it over unless someone else volunteers.

[xingyaoww]

you will see the parent agent die while its child still running.

On second thought, we do hope the parent agent dies when it exceeds the (global)_max_iteration. Maybe i can just move this code block back for this PR, and wait for the gloabl_max_iteration fix.

[li-boxuan]

Fix is coming! #2990

[xingyaoww]

I think this is still not fixed yet :( -

We probably need the headless_mode change from https://github.com/OpenDevin/OpenDevin/pull/2994/files#diff-37dff7c7cda815bf00598e78acaaebb91df0bf68b820bdca78a86b89f67dfc9b to fix it:

[li-boxuan]

Right, a new problem arises after fixing the old problem lol