You will need
- Claude Code (included in Claude paid plans) or another agent tool that supports subagents
- Web access for the agent: built-in web search, or a search/scrape API such as Firecrawl (free tier: 1,000 credits a month)
- A research question with a clear scope and a decision it supports
You will end up with
- A one-page brief where every claim links to a source
- Each claim marked supported, unsupported or unclear by a separate checker
- A list of open questions to follow up
Write the research question
One sentence on what you need to know and for which decision. Add what is out of scope and how recent the sources must be. Anthropic's own research system starts by having a lead agent break the question into subtasks, each with a clear objective.
Sources: Anthropic Engineering: How we built our multi-agent research system
Decide if you need several agents at all
Multi-agent runs cost far more: Anthropic reports multi-agent systems use about 15 times more tokens than a chat. For a narrow question, one agent with search is enough. Use this playbook for broad questions with many independent parts.
Sources: Anthropic Engineering: multi-agent research system (token cost) · Anthropic: Building effective agents (keep it simple)
Create three subagents
In Claude Code, define a gatherer, a writer and a checker as subagents. Each runs in its own context window with its own prompt and tool access. Give only the gatherer web tools.
Sources: Claude Code docs: Subagents · Firecrawl docs: Search · Firecrawl docs: Scrape
Prompt Gatherer: find [N] recent, primary sources on [QUESTION]. For each, return the URL, publish date, and one line on why it matters. Skip sources older than [DATE]. Writer: write a one-page brief using only these sources. Put the source URL after every claim. Checker: for each claim, open the cited URL and mark it supported, unsupported or unclear, quoting the sentence that supports it.
Run the checker on every claim
A separate checking pass is the evaluator-optimizer pattern: one model writes, another evaluates. Anthropic's research system also runs a dedicated citation step so every claim is attributed to a source.
Sources: Anthropic: Building effective agents · Anthropic Engineering: multi-agent research system (CitationAgent)
Read the checked brief
Keep only claims marked supported. Open the unclear ones yourself before you act on them. Web pages can contain instructions aimed at agents (prompt injection), so treat fetched text as data.
Sources: Anthropic: Mitigating prompt injections in browser use
Human checkpoint. A person approves this step before the agent goes further.