Creativity · Agent Protocol
Deep Research Agent Pattern
'Deep research' has become a shared product category across frontier labs: Perplexity's Deep Research, Google Gemini's Deep Research, OpenAI's ChatGPT Deep Research, and Anthropic's Claude Research. Under the hood they share a pattern — a lead agent decomposes the question into sub-queries, dispatches parallel sub-agents to search and read, and a synthesiser stitches the findings into a cited report. It's the flagship mainstream use case for multi-agent orchestration.
Protocol facts
- Sponsor
- shared pattern (Anthropic, OpenAI, Google, Perplexity)
- Status
- stable
- Spec
- https://www.anthropic.com/engineering/multi-agent-research-system
- Interop with
- MCP, A2A, search APIs, browser tools
Frequently asked questions
Why parallel sub-agents instead of one long chain?
A single sequential chain is limited by its own bandwidth. Parallel sub-agents cover more of the search space concurrently, and narrow per-agent context keeps each sub-search focused — both factors Anthropic credits for the quality lift their research agent shows.
Is deep research just 'search with citations'?
It's more than that — the agent reformulates the question, decides which sub-questions to pursue, reads full documents rather than snippets, and synthesises across sources. The citations are the tip of the iceberg.
Why does this use more tokens than single-agent search?
Because it reads more. A deep-research run often traverses dozens of pages; the cost scales with breadth. The trade-off is depth — you're paying for research quality, not just an answer.
Sources
- Anthropic — How we built our multi-agent research system — accessed 2026-04-20