In 2025, Anthropic gave its Claude chatbot a feature called Research. Ask it a big question and it doesn’t just search once. One model, the lead, plans the job, then sends several helper models out to search the web at the same time, each in its own lane. The lead gathers what they bring back and writes the answer.
Getting there took a while, and the engineers later wrote up what went wrong.
The early lead agent managed like someone who had never managed before. It spun up 50 helpers to answer simple questions. Some helpers kept searching for sources that didn’t exist. Others distracted each other with a steady stream of updates.
Its instructions were the weakest part. The lead would hand out a one-line task such as “research the semiconductor shortage.” Left to interpret that, one helper studied the 2021 car chip crisis, while two others both dug into 2025 supply chains, running the same searches as each other. No one had divided the work.
The team started with the lead’s instructions for handing out work. Every helper now had to receive four things: an objective, a format for its answer, guidance on which tools and sources to use, and clear boundaries for its task. The lead also got rules of thumb for sizing the team: one agent and three to ten tool calls for a simple fact, two to four helpers for a comparison, and, for complex research with clearly divided responsibilities, possibly more than ten.
Along with work on tools and evaluation, that changed the system. On Anthropic’s internal research tests, the finished team, led by Claude Opus 4 with Claude Sonnet 4 helpers, beat a single Claude Opus 4 by 90.2%. Running helpers and their searches in parallel cut research time by up to 90% for complex questions.
It wasn’t free. A multi-agent run burned about 15 times as many tokens as an ordinary chat. And the engineers were blunt that some work doesn’t suit a team at all: tasks where every agent needs the same context, or where the pieces depend on each other. Most coding, they wrote, has fewer truly parallel parts than research.