• Deep research has become one of the most popular agent applications
• Major players with deep research products:
  • OpenAI
  • Anthropic
  • Perplexity
  • Google
• All produce comprehensive reports using various sources of context
• Many open source implementations available

Research is an open-ended task; the best strategy to answer a user request can't be easily known in advance.

Validation questions require iterative deep research where source quality matters more than breadth.

Agents are well suited to research because they can flexibly apply different strategies, using intermediate results to guide exploration.

• Scope – clarify research scope
• Research – perform research
• Write – produce the final report

User Clarification

Users rarely provide sufficient context in a research request. We use a chat model to ask for additional context if necessary.

Brief Generation

We translate the verbose chat interaction into a comprehensive, yet focused research brief that serves as our north star for success.

Research Supervisor

The supervisor delegates research tasks to an appropriate number of sub-agents. It determines if the research brief can be broken down into independent sub-topics and delegates to sub-agents with isolated context windows.

Research Sub-Agents

Each sub-agent focuses on a specific topic and conducts research as a tool-calling loop, using search tools and/or MCP tools configured by the user.

The goal of report writing is to fulfill the request in the research brief using the gathered context from sub-agents.

When the supervisor deems that the gathered findings are sufficient to address the request in the research brief, we move ahead to write the report.

To write the report, we provide an LLM with the research brief and all of the research findings returned by sub-agents. This final LLM call produces an output in one-shot, steered by the brief and answered with the research findings.

Only use multi-agent for easily parallelized tasks

Multi vs. single-agent is an important design consideration. Cognition has argued against multi-agent because sub-agents working in parallel can be difficult to coordinate.

Earlier versions of our research agent wrote sections of the final report in parallel with sub-agents. It was fast, but the reports were disjoint because the section-writing agents were not well coordinated.

Multi-agent is useful for isolating context across sub-research topics

Single agent response quality suffers if the request has multiple sub-topics. The intuition is straightforward: a single context window needs to store and reason about tool feedback across all sub-topics.

Our single agent implementation used its search tool to send separate queries about each frontier lab at the same time, but had to juggle context from three independent threads.

Multi-agent supervisor enables the system to tune to required research depth

Users do not want simple requests to take 10+ minutes. But, there are some requests that require research with higher token utilization and latency.

The supervisor can handle both cases by selectively spawning sub-agents to tune the level of research depth needed for a request. The supervisor is prompted with heuristics to reason about when research should be parallelized, and when a single thread of research is sufficient.

Context Engineering is important to mitigate token bloat and steer behavior

Research is a token-heavy task. Anthropic reported that their multi-agent system used 15x more tokens than a typical chat application!

We used context engineering to mitigate this:

• We compress the chat history into a research brief
• Sub-agents prune their research findings to remove irrelevant tokens

Without sufficient context engineering, our agent was prone to running into context window limits from long, raw tool-call results.

Open deep research is a living project and we have several ideas we want to try. These are some of the open questions that we're thinking about:

• What is the best way to handle token-heavy tool responses, and what is the best way to filter out irrelevant context to reduce unnecessary token expenditure?
• Are there any evaluations worth running in the hot path of the agent to ensure high quality responses?
• Deep research reports are valuable and relatively expensive to create, can we store this work and leverage these in the future with long-term memory?