fix(deep_research): resolve DeepResearchAgent returning None (fixes #1770)

[aviraldua93]

Fix DeepResearchAgent returning None

Fixes #1770

Problem

DeepResearchAgent returns None instead of the researched answer. The root cause is in DeepResearchTool, which uses three nested initiate_chat() calls that terminate via tool execution (confirm_summary, generate_subquestions, confirm_answer).

When a chat terminates via tool call, the default _last_msg_as_summary method extracts content from the last sender→recipient message — but that message is a tool_calls message with content=None. This causes result.summary to return "" for all three inner chats, propagating up as None.

Root Cause Trace

1. delegate_research_task() → critic_agent.initiate_chat(summarizer_agent, ...) — terminates when confirm_summary tool is called
2. _last_msg_as_summary() (conversable_agent.py L2159) → recipient.last_message(sender)["content"] → None (tool_call message)
3. result.summary → "" → returned up the chain → DeepResearchAgent returns None

Same pattern affects all 3 inner chats: confirm_summary, generate_subquestions, and confirm_answer.

Fix

Added _extract_last_result_with_prefix() static method to DeepResearchTool that:

1. Walks chat_result.chat_history in reverse
2. Finds the first message whose content starts with the expected prefix (e.g., "Answer confirmed:", "Subquestions answered:", "Subquestions generated:")
3. Falls back to result.summary if no match is found

Replaced all 3 bare return result.summary calls with this helper, passing the appropriate prefix for each chat stage.

Tests

Added 5 unit tests in TestExtractLastResultWithPrefix:

• Extracts answer from tool result message in chat history
• Returns last match when multiple prefixed messages exist
• Falls back to summary when no prefix match found
• Handles None content gracefully (no crash)
• Handles empty chat history

All tests pass without requiring API keys or optional dependencies.

Changes

• autogen/tools/experimental/deep_research/deep_research.py — Added helper method + replaced 3 return statements
• test/tools/experimental/deep_research/test_deep_research.py — Added 5 regression tests

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[amabito]

deep_research.py:83-88 — list-type content silently skipped

isinstance(content, str) skips non-string content without a trace. Multimodal content would be a quiet no-op here — maybe a logger.debug on the else branch?

deep_research.py:135-142 — silent_override=True not explained

Four silent_override=True additions aren't in the PR description. Suppressing intermediate tool output? Curious how this connects to the None fix — or is it a separate cleanup?

[aviraldua93]

deep_research.py:83-88 — list-type content silently skipped

isinstance(content, str) skips non-string content without a trace. Multimodal content would be a quiet no-op here — maybe a logger.debug on the else branch?

deep_research.py:135-142 — silent_override=True not explained

Four silent_override=True additions aren't in the PR description. Suppressing intermediate tool output? Curious how this connects to the None fix — or is it a separate cleanup?

Thanks @amabito Thanks for the careful review — both points are valid.

silent_override=True: You're right to call this out — it was unrelated to the None fix and shouldn't have been in this PR. Removed all 4 instances; the
register calls now match the original code.

List-type content: The three tool functions (confirm_summary, confirm_answer, generate_subquestions) all explicitly return str, so list content won't
appear in practice here. That said, defensive logging is cheap — I've added a logger.debug on the elif isinstance(content, list) branch so it's not a
silent no-op. Also hardened the fallback to chat_result.summary or "" to match the -> str return type annotation.

Updated in 29c563c. Thanks again for the feedback!