Researchers split big web searches across many AI agents, borrowing an old computing trick
A new arXiv paper applies the MapReduce idea to AI agents, so many small agents each take one piece of a broad search and a final step merges the results.
The short version
- A-MapReduce is built for 'wide search': collecting many items from many sources, not digging deep into one topic.
- It splits the job into pieces, runs agents in parallel, then merges the results in a structured way.
- It keeps a memory of past experience to improve how future tasks get divided up.
- Tested on five benchmarks, with average Item F1 gains of 5.11% - 17.50% over strong baselines.
- The authors say it also cut running time versus representative multi-agent baselines, and they released the code.
Why it matters
Many real research requests, like comparing every competitor's pricing, are wide rather than deep, and a single assistant working through them one at a time tends to bog down. This paper is evidence that splitting such work into parallel pieces can help. It is still a preprint from the authors themselves, so treat the results as promising rather than settled.
Wide versus deep
The paper starts from a distinction between two kinds of research tasks. Deep research follows a trail step by step, with each answer shaping the next question. Wide search is different: it means gathering a large number of items across many sources.
According to the abstract on arxiv.org, today's AI agent systems are mostly designed for the first kind. When they face wide tasks, they struggle with the sheer size of the goal and with slow, long-running execution.
Borrowing MapReduce
MapReduce is an older computing idea: break a large job into small pieces, process them side by side, then combine the outputs. A-MapReduce applies that pattern to teams of AI agents.
The system decomposes a task according to what it needs, sends the pieces to agents that work in parallel, and then aggregates the results in a structured form. It also keeps an experiential memory, meaning a record of past runs, which it uses to improve how later tasks are divided and recombined.
What the authors report
The team tested the framework on five agentic benchmarks, which are standardized tests for agent systems. They say it reached state-of-the-art results on two of them, WideSearch and DeepWideSearch.
Against strong baselines built on OpenAI o3 or Gemini 2.5 Pro, average Item F1 scores rose by 5.11% - 17.50%. Item F1 is the accuracy score the paper uses for the items an agent finds. The authors also claim a better cost-to-performance balance and a shorter running time than representative multi-agent baselines.
Caveats
These are the authors' own claims, and the paper is a preprint, so independent checks are still to come. The abstract does not say how the system behaves on tasks outside these benchmarks. The code is public on GitHub, which lets others test it.
What to watch
- Whether independent teams reproduce the reported gains using the released code.
- Whether the approach holds up on real-world wide searches beyond the five benchmarks.
Try it
Next time you ask an assistant for a long list, such as pricing from ten competitors, ask one small question per item in separate chats, then request a combined table.
Sources and links
- Reportingarxiv.org: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by arxiv.org · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.