[BUG][Multi-Agent Structures][Eleven structures never reset their Conversation per task, and two fan run() across threads sharing one]

[kyegomez]

Summary

Follow-up to #2041. A sweep found eleven structures that build their Conversation in __init__ and never reset it in run(). A second run() on the same instance therefore serves the previous task's history. Two of them additionally fan self.run across threads sharing that one object, which is a data race rather than just a leak.

The two races

hybrid_hiearchical_peer_swarm.py — conversation built at :121, and :212:

run_concurrently(self.run, tasks, ...)

N threads call self.conversation.add() on one object, and each returns history_output_formatter(self.conversation, ...) — so every caller receives all tasks' messages, not its own.

multi_agent_router.py — conversation built at :185, and :485:

executor.submit(self.route_task, task)

Same shape.

The leaks

Sequential reuse, so no race, but task N sees task N-1's history:

File	Conversation built	Batch entry point
advisor_swarm.py	:118	:267 batched_run
llm_council.py	:318	:534 batched_run
round_robin.py	:134	:322 batched_run
debate_with_judge.py	:182	:555 batched_run
auction_swarm.py	:286	:470 batch_run
council_as_judge.py	:289	repeated run()
heavy_swarm.py	:171	repeated run()
planner_worker_swarm.py	:586	repeated run()
planner_generator_evaluator.py	:261	repeated run()

Non-Conversation state with the same problem: spreadsheet_swarm.py:91-93 (self.outputs, self.tasks_completed, self.agent_tasks) and hierarchical_structured_communication_framework.py:1168 (conversation_history, evaluation_results, intermediate_outputs).

Fix

Uniform: reset at the top of run()

self.conversation = Conversation(...)

or accept an injected conversation the way planner_worker_swarm.WorkerPool already does (:369).

For the two races, resetting is not sufficient — a shared instance cannot be safely fanned across threads at all. Either give each task its own conversation (the _clone_for_task approach AgentRearrange.batch_run uses, agent_rearrange.py:1032) or serialize.

Already correct

ma_blocks.py:57, deep_discussion.py:39, and multi_agent_debates.py:49/:121 build a fresh Conversation() inside the function, so they are correct by construction. Worth using them as the reference shape.

Found at 3e89f27b (v14.0.2).