Add AG2 integration for multi-agent tracing

[faridun-ag2]

Summary

• Adds deepeval/integrations/ag2/ with instrument_ag2() that patches AG2's ConversableAgent to automatically capture AgentSpan, LlmSpan, and ToolSpan for DeepEval evaluation
• Follows the same pattern as the existing CrewAI and LangChain integrations (monkey-patching + Observer context managers)
• Includes usage example in examples/tracing/ag2_tracing.py

What it captures

AG2 method	DeepEval span	Data captured
generate_reply / a_generate_reply	AgentSpan	Agent name, reply output
OpenAIWrapper.create	LlmSpan	Model name, messages, token counts
execute_function / a_execute_function	ToolSpan	Function name, arguments, result

Usage

from deepeval.integrations.ag2 import instrument_ag2
instrument_ag2()

# Run AG2 agents as usual — traces captured automatically
executor.run(assistant, message="...").process()

Why AG2

AG2 is an open-source multi-agent orchestration framework with 500K+ monthly PyPI downloads. It's a natural addition alongside the existing CrewAI, LangChain, and LlamaIndex integrations.

Files

deepeval/integrations/ag2/
├── __init__.py      # Public exports
├── handler.py       # instrument_ag2() / reset_ag2_instrumentation()
├── wrapper.py       # Method wrapping (generate_reply, execute_function, OpenAIWrapper.create)
examples/tracing/
├── ag2_tracing.py   # Usage example

No changes to existing code.

Test plan

• Import verification: from deepeval.integrations.ag2 import instrument_ag2 works
• Lifecycle: instrument → verify patched → reset → verify restored
• AgentSpan: generate_reply creates span with correct agent name
• ToolSpan: execute_function creates span with function name, args, output
• Nested spans: tool span appears as child of agent span
• Error handling: failed tools get ERRORED status
• E2E: full conversation with gpt-4o-mini produces traces
• Linting: black + ruff pass cleanly

[vercel]

@faridun-ag2 is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[faridun-ag2]

Hi @penguine-ip @kritinv!
We'd love your review on this when you get a chance!

This PR adds an AG2 multi-agent tracing integration following the same pattern as the existing CrewAI and LangChain integrations. All new files, no changes to existing code.

Note on CI: The test_core job fails on forked PRs due to a pre-existing issue — the ignore path for test_dataset_iterator.py is missing the test_integration/ subdirectory:

--ignore=tests/test_core/test_tracing/test_dataset_iterator.py            # current (wrong)
--ignore=tests/test_core/test_tracing/test_integration/test_dataset_iterator.py  # correct

This causes the test to run without an OPENAI_API_KEY and fail. The black formatting failures (20 files) were also pre-existing on main — I've included a fix for those in a separate commit.

Happy to address any feedback!