How do you design a deep-research agent?

In this article

1. Before you start
2. Why interviewers ask this
3. Deep-research agent from first principles
4. Worked example
5. Trade-offs and edge cases
6. How to answer it in an interview
7. Common mistakes
8. Follow-up questions
9. Sources

Before you start

This page builds on AI agent architecture and prompting.

Why interviewers ask this

The prompt "design a deep-research agent" is reported by candidates in interviews at AI labs and large tech companies, in several wordings: answer a complex question by searching and synthesizing many sources, or build a multi-agent system that produces a cited report. The consequence of a vague answer is a long list of follow-ups, because every box you draw needs a reason.

Interviewers use it to see whether you can split a large job across workers, control cost, and keep claims tied to evidence. A weak answer draws "search, then summarize" in a single loop. A strong answer separates planning, searching, writing and citing, and says why. A Staff-level answer adds the budget, the failure modes (runaway agents, endless searching) and how you would evaluate the result. The broader AI agent system design answer covers the generic loop, tools and memory this page builds on.

Deep-research agent from first principles

Deep-research agent

A deep-research agent is a system built from large language model (LLM) calls that turns an open-ended question into a complete workflow of planning, delegation, synthesis and citation, running under an explicit research budget controlling agent count, tool usage and reasoning depth.

Subagent

A subagent is a worker agent started by the lead agent for one sub-task, with its own instructions, tools and context window, that returns only distilled findings to the lead.

In plain words

Picture a newsroom editor with a big story. The editor drafts a plan. Reporters then go off to separate angles. The editor collects their notes and writes the piece. A fact-checker attaches the sources. The reporters never read each other's notebooks.

The analogy breaks in one place. A human editor knows when a story needs five reporters. A model has to be told, so the rules for sizing the team live in its prompt.

In one precise sentence, a deep-research agent is a workflow of agents. An orchestration agent coordinates specialized worker agents. They operate in parallel on different aspects of a research task (arXiv survey, late 2025). Anthropic describes its own version as an agent that plans a research process and then creates parallel agents that search simultaneously (June 2025).

Step 1: The lead agent plans and saves the plan

The lead agent reads the question, thinks through an approach and writes a plan. In Anthropic's design (June 2025) the lead saves that plan to memory, because if the context window grows past 200,000 tokens it is truncated and the plan would be lost. A context window is the amount of text a model can read at once, so anything that must survive a long run belongs in external memory, not only in the conversation.

Input: the user's question. Output: a written plan that lists the facets to investigate. You now know where the work starts: planning is a separate step, and its output is stored outside the window. The risk of a growing window is covered in context rot.

Step 2: The lead scales effort to the question

The lead agent decides how many subagents a question deserves, using rules written into its prompt. Anthropic reports that early agents spawned 50 subagents for simple queries, searched endlessly for sources that did not exist and distracted each other. Its fix was effort rules: simple fact-finding needs 1 agent with 3 to 10 tool calls, direct comparisons 2 to 4 subagents with 10 to 15 calls each, and complex research more than 10 subagents.

The survey describes the same idea as an explicit research budget on agent count, tool usage and reasoning depth. Anthropic's stated reason for putting the rules in the prompt is that agents struggle to judge appropriate effort for different tasks. Input: the plan. Output: a list of subagent briefs. Each brief needs an objective, an output format, guidance on tools and sources, and clear task boundaries, a list also quoted by Simon Willison. This is the answer to the trap in the lead: a budget plus effort rules in the prompt.

Step 3: Subagents search in their own contexts

Each subagent runs its own search loop and returns distilled findings, not raw pages. Anthropic describes search as compression: subagents work in parallel with their own context windows, exploring different aspects at once before condensing the most important tokens for the lead. A separate window per subagent is how more material fits than one window can hold.

Speed comes from two kinds of parallelism: the lead starts 3 to 5 subagents in parallel, and each subagent uses 3 or more tools at the same time. Anthropic reports that these changes cut research time by up to 90% for complex queries. Tools matter too: poor tool descriptions mislead agents, which is why tool interfaces such as the Model Context Protocol need clear descriptions.

Step 4: The lead synthesizes and a citation agent attributes claims

The lead agent merges the subagents' findings and decides whether more research is needed; if so it creates more subagents or refines its strategy. When enough is gathered, the system leaves the research loop and passes everything to a citation agent, which identifies specific locations for citations in the documents and the draft report.

Input: distilled findings. Output: a report whose claims point to sources. Keeping citation as its own step is a design choice: the writer focuses on the answer, and a separate pass ties each claim to a passage. Sources are not attributed automatically just because the agent saw them. A human check on high-stakes reports fits here; see human in the loop.

1 · Plan2 · Scale effort3 · Search4 · Synthesize and citeReplay

Step-by-step description

1. The lead agent plans and saves the plan.
2. Effort rules size the team.
3. Subagents search in their own contexts.
4. The lead writes, a citation agent attributes.

• The lead agent plans the research and saves the plan to memory.
• Effort rules decide how many subagents a question needs.
• Each subagent searches in its own context window and returns distilled findings.
• The lead synthesizes the findings and a citation agent attributes the claims.

Worked example

This budget shows why the design helps and what it costs. All inputs are illustrative: a subagent reads 50,000 tokens while searching and returns a 2,000-token summary, the lead's plan is 1,500 tokens, each subagent makes 8 searches of 15 seconds, and the lead adds 30 seconds.¹ The 200,000-token window and the 4 times and 15 times token multiples come from Anthropic's article (June 2025).

With 5 subagents the team reads 250,000 tokens, more than one 200,000-token window holds, yet the lead's own window carries only 11,500 tokens (plan plus five summaries).¹ With 10 subagents the team reads 500,000 tokens and the lead holds 21,500.¹ A single agent with no lead, running every search in turn, needs 600 seconds for five subagents' worth of searches, against 150 seconds for the lead and five parallel subagents, a 75% saving; for ten it is 1,200 against 150.¹ Illustrative single-agent time is 120 seconds for one subagent's searches and 1,200 seconds for ten, while time with a lead and parallel subagents is 150 seconds whatever the number of subagents.

For one narrow question, a single agent (120 s) beats a lead with one subagent (150 s): the orchestration adds overhead and tokens and buys no parallelism, so a narrow question should not use a team.¹ The cost side: for an illustrative 5,000-token chat turn, an agent at about 4 times chat uses 20,000 tokens and a multi-agent system at about 15 times uses 75,000.¹

Text description of the diagram

• Illustrative single-agent time is 120 seconds for one subagent's searches and 1,200 seconds for ten.
• Illustrative time with a lead and parallel subagents is 150 seconds whatever the number of subagents.

Trade-offs and edge cases

Multi-agent research buys breadth and speed with tokens. Anthropic's data show agents use about 4 times more tokens than chat and multi-agent systems about 15 times, so the task must be valuable enough to pay for it. In an internal research evaluation, a Claude Opus 4 lead with Claude Sonnet 4 subagents outperformed a single Opus 4 agent by 90.2%. On the BrowseComp benchmark, token usage alone explained 80% of the variance, which suggests the gain comes largely from spending more tokens across separate windows.

Text description of the diagram

• A single agent uses about 4 times the tokens of chat and a multi-agent system about 15 times.
• Multi-agent systems fit broad, parallel work and fit poorly when all agents need the same context.

Edge cases: Anthropic says domains where all agents must share the same context or have many dependencies fit poorly. Its lead agents run subagents synchronously, waiting for each set to finish, which simplifies coordination but creates a bottleneck. Production needs deterministic safeguards such as retry logic and checkpoints, and rainbow deployments (shifting traffic gradually between versions) so updates do not break running agents. Anthropic embedded scaling rules in the prompts because agents struggle to judge appropriate effort for different tasks. For why long runs break, read why agents fail on long tasks.

How to answer it in an interview

The spoken structure: name the pattern, walk plan, delegate, search, synthesize, cite, then close on the budget and the cost trade-off.

Common mistakes

• "More agents always give a better report." (wrong) Early agents spawned dozens of subagents for simple queries. (right) "I scale the number of subagents to the question's complexity and cap it with a budget."
• "Deep research is one big prompt." (wrong) It is a workflow of stages. (right) "Planning, delegation, synthesis and citation are separate steps."
• "The agent saw the page, so the citation is automatic." (wrong) Attribution needs its own pass. (right) "A citation step matches claims to passages in the sources."
• "Every task needs multiple agents." (wrong) Dependent, shared-context work fits poorly. (right) "I use multi-agent for broad, parallel facets and one agent otherwise."

Follow-up questions

How do you create a system where agents collaborate to produce a research report with citations?

Give the lead agent the plan and the final merge, give each subagent a brief with an objective, output format, tool guidance and boundaries, and end with a citation agent. Agents collaborate through written briefs and returned findings, not shared notebooks.

How would you design a semantic cache for LLM responses that reduces cost and latency?

A semantic cache is a store that returns a saved answer when a new query's embedding is close enough to an earlier one, so the model is not called again. In this design the cache sits in front of the subagents' search calls and the planning step, so repeated queries reuse earlier search results or a saved plan, not only final reports. Too loose a similarity threshold causes false hits, where a different question gets a wrong saved answer, so set it carefully. Give each entry a time-to-live (TTL) by source, short for fast-changing topics, and be cautious caching cited research reports, whose sources may change.

How do you evaluate a research agent?

Score the final report for correctness and for whether each claim matches its cited passage, and also inspect the trajectory of tool calls. See evaluating agents for trajectory and reliability checks.

When should you not use multiple agents?

Skip them for narrow questions, tasks with many dependencies between steps, and cases where all agents need the same context; multi-agent systems covers that decision. The token cost of roughly 15 times chat is not worth it there.

What happens when a subagent searches forever?

Cap tool calls per subagent and total agents in the budget, and prompt it to stop when sources do not exist. Anthropic reports endless searching for nonexistent sources as an early failure. More agent questions are in agentic AI interview questions.

Sources

1. How we built our multi-agent research system — Anthropic Engineering, 13 Jun 2025
2. Deep Research: A Systematic Survey (arXiv 2512.02038) — arXiv, 1 Dec 2025
3. Anthropic: How we built our multi-agent research system (commentary) — Simon Willison's Weblog, 14 Jun 2025