---
title: "McGill's Mujica-MyGO Splits AI Search Agents to Beat Context Overload"
subtitle: "A new arXiv paper proposes Mujica-MyGO, splitting multi-hop retrieval into a planner-worker DAG and training it with a stripped-down policy gradient method."
url: "https://alphasignal.ai/news/mcgill-s-mujica-mygo-splits-ai-search-agents-to-beat-context-overload"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-14T00:00:00.000Z"
updated: "2026-08-24T02:22:57.270Z"
type: "Paper"
topics:
  - "Agents"
  - "Retrieval"
subtopics:
  - "Multi Agent"
  - "Rag"
upvotes: 2358
source_url: "https://arxiv.org/abs/2505.17086"
comments: 0
truncated: true
---

# McGill's Mujica-MyGO Splits AI Search Agents to Beat Context Overload

> A new arXiv paper proposes Mujica-MyGO, splitting multi-hop retrieval into a planner-worker DAG and training it with a stripped-down policy gradient method.

Multi-turn retrieval-augmented generation has a scaling problem that gets uglier the deeper the model reasons. Every search call, every retrieved passage, every intermediate rationale gets stuffed back into the prompt, and the context balloons until the model loses track of what it was doing. A new preprint from researchers at McGill, Mila, the Chinese University of Hong Kong, Tianjin University, and Huawei Noah's Ark Lab proposes a pragmatic fix: split the problem across two cooperating agents, then train them with the simplest reinforcement learning update the authors could get to converge.

The framework is called [Mujica-MyGO](https://arxiv.org/abs/2505.17086), and it targets a bottleneck anyone building agentic RAG has hit. Modern RAG pipelines interface with search engines over multiple turns for complex reasoning tasks, and those interactions produce long intermediate contexts, with context length growing exponentially with exploration depth. That collides with a well-known weakness of LLMs: their difficulty leveraging information from long contexts. The issue compounds in systems that depend on in-context learning, where few-shot demonstrations must also live in the prompt.

## Divide, conquer, then answer

The Mujica half of the system is a multi-agent workflow that leans on a classical divide-and-conquer trick. It consists of two modules, a planner and a worker, each of which serves as a user and an assistant for the other. The planner extracts subquestions, organizes them into a directed acyclic graph (DAG), and iteratively calls the worker to solve each one.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning (arXiv abstract)](https://arxiv.org/abs/2505.17086)
- [Advancing Multi-Agent RAG Systems with Minimalist Reinforcement Learning (arXiv HTML v4)](https://arxiv.org/html/2505.17086)
- [Reinforcing Question Answering Agents with Minimalist Policy Gradient Optimization (v2 HTML)](https://arxiv.org/html/2505.17086v2)
- [ResearchGate: Reinforcing Question Answering Agents with Minimalist Policy Gradient Optimization](https://www.researchgate.net/publication/392085894_Reinforcing_Question_Answering_Agents_with_Minimalist_Policy_Gradient_Optimization)