Deep Research Agents: Cited Answers on Autopilot

Most "AI research" is a single prompt and a wall of unsourced text you still have to fact-check by hand. A deep research agent works differently: it plans the question, searches multiple sources, reads them, and hands you an answer with citations you can click.

What an AI deep research agent actually does

A deep research agent is not a one-shot chatbot reply. It is a multi-step process: it breaks your question into sub-questions, runs web searches, opens and reads the results, cross-checks what it finds, and then writes a structured answer with inline citations pointing back to each source.

In AgentWorks, Deep Research is one of the built-in tools inside multi-LLM chat, alongside web search, image generation, code execution, and your company knowledge base. You ask a question the way you'd brief a junior analyst — "compare the EU market entry rules for X across these three countries" — and the agent returns a sourced write-up instead of a paragraph of unverifiable claims. Every factual claim is tied to where it came from, so you review the reasoning, not just the conclusion.

Because it runs on the same platform as everything else, you can start a research run and then keep working in the same conversation: refine the scope, switch models mid-thread, or push the output straight into a document.

Why citations and multi-step reasoning matter

The value of research is only as good as your ability to trust it. Unsourced AI text forces you to redo the work — you either verify every sentence or you take a risk. Citations remove that tax: you can open the source, confirm the number, and move on.

Multi-step reasoning matters for the same reason. A single search rarely answers a real question. Comparing vendors, tracking a regulation across markets, or summarising a research field all require several passes: find the candidates, read each one, reconcile the differences. A deep research agent does those passes for you and shows its trail.

AgentWorks pairs this with a grounding rule that applies across the platform: when an answer isn't supported by the sources or your knowledge base, the agent says "I don't know" rather than inventing one. That honesty is the difference between a tool you can put in front of a client and one you can't.

Run it on your own documents, not just the web

Public web research is half the picture. The other half is your own material — reports, contracts, past analyses, internal wikis. AgentWorks lets a deep research agent draw on both.

Upload PDF, DOCX, TXT, or CSV files, or connect sources like Notion, Confluence, and URLs, and the platform indexes them with pgvector so the agent can retrieve and cite the exact passages it used. Sensitive data is masked at the gateway before any model sees it, and you can read more about how retrieval-augmented generation is wired up. The result is research that blends live external sources with your organisation's own knowledge, each answer traceable to a specific document.

This is where a research agent stops being a novelty and becomes part of the workflow: it answers questions the way your best analyst would, using both what's on the web and what's in your files.

Schedule research and export the results

A one-off answer is useful. A standing research process is a competitive edge. On the Pro plan and above, you can schedule agents to run daily, weekly, or monthly — or trigger them from a webhook when something changes upstream.

Set up a weekly competitor scan, a Monday market briefing, or a monthly regulatory digest, and the output lands without anyone kicking it off. Chain research into a longer multi-agent pipeline — research, then draft, then review, then publish — where each step is logged and carries its own risk class.

When the research is done, you export it. AgentWorks builds Word, PowerPoint, Excel, and PDF files in a live canvas you can open directly in Google Drive or OneDrive, so a research run becomes a finished briefing document rather than text you have to reformat. The same integrations that feed the agent — Slack, Teams, Gmail, SharePoint, and more — can also deliver the result to where your team already works.

Pick the right model, pay only for what you use

Deep research can be heavy work, so cost control matters. AgentWorks runs on multiple model families — GPT-5 and GPT-5 mini, Claude Opus, Sonnet, and Haiku, Gemini Pro and Flash with up to a 1M-token context, and Mistral Large — and you can see the full model line-up.

The AUTO router sends each message to the cheapest model that can handle it, so a simple sub-query doesn't pay Opus prices. Tokens are billed at cost plus 10% from a single transparent euro wallet, with live per-run spend and budgets you can set per organisation, team, or user. You see exactly what a research run cost before it surprises you on an invoice, and you can start on the Free plan with a €5 one-time credit to try it.

Governance built for European teams

Research often touches regulated or confidential material, so the controls around it matter as much as the output. AgentWorks is EU-native, built in the Netherlands, and designed to be EU AI Act-ready — meaning each agent carries a risk classification and the platform gives you the controls to operate responsibly, not a blanket compliance promise (your actual risk depends on your use case).

Every research run writes to an immutable, append-only audit trail you can export as CSV or JSON, state-changing actions can require human approval, and data stays in the EU with model endpoints offered on European infrastructure. Model contracts are no-training and zero-retention, and a DPA is available on request. You can review the full picture on the trust and compliance pages.

Summary: An AI deep research agent runs multi-step, multi-source research and returns cited, verifiable answers. AgentWorks lets you point it at the web and your own documents, schedule it to run on autopilot, export the results to Word, PDF, or Excel, and keep every run under EU-native governance — with the AUTO router keeping costs at cost plus 10%.

Frequently asked questions

How is a deep research agent different from a normal AI chat answer?

A normal chat answer is a single pass that often gives you unsourced text. A deep research agent plans the question, runs several searches, reads the results, and returns a structured answer with inline citations. You can trace every claim to its source instead of taking it on faith.

Can I run research on my own company documents?

Yes. You can upload PDF, DOCX, TXT, and CSV files or connect sources like Notion and Confluence, and the agent retrieves and cites the exact passages it used. If an answer isn't supported by your knowledge base, the agent says "I don't know" rather than guessing, and sensitive data is masked before any model sees it.

Can research runs be scheduled and automated?

On the Pro plan and above you can schedule agents to run daily, weekly, or monthly, or trigger them from a webhook. Research can also be chained into a multi-agent pipeline and the output exported to Word, PowerPoint, Excel, or PDF, then delivered through integrations like Slack, Teams, or email.