Every major AI company now ships a "deep research" feature. Google has it. OpenAI has it. Perplexity has it. They're all impressive in demos — ask a complex question, wait a few minutes, get a multi-page report with sources.
The problem isn't the quality. The problem is that none of these tools were built for your agent to use.
They live inside chat interfaces. They produce reports formatted for human reading, not structured data for downstream processing. And if your agent can't call a tool programmatically, that tool might as well not exist for your workflow.
What deep research actually does differently
Regular search — even grounded search — answers one question in one pass. You ask, it retrieves, it synthesizes.
Deep research breaks a complex question into sub-questions, runs multiple rounds of search, cross-references conflicting sources, and compiles findings into a structured report. It's the difference between "what's Acme's pricing" and "analyze the competitive landscape for enterprise AI search tools, including pricing, differentiation, and developer sentiment."
The output isn't a paragraph. It's 20-100+ sources synthesized into something closer to an analyst report. The latency is accordingly longer — 2 to 15 minutes rather than seconds. The cost is higher — $0.50 to $5+ per report rather than fractions of a cent.
The tools, ranked by whether your agent can actually use them
AnyCap Deep Research is the only one purpose-built for agents. You install it as a skill (claude mcp add anycap-cli-nightly), and your agent invokes it like any other tool. The output is structured — JSON with sections, citations, and confidence scores, not just a text report. Your agent can parse it, filter it, and feed it into the next step of a workflow:
anycap research \
--query "AI agent capability runtime market Q2 2026" \
--depth comprehensive --output market-analysis.md
Google Gemini Deep Research produces good reports. It has Google's search index behind it, which matters for retrieval quality. But the API output is formatted text — no structured citations, no JSON sections. Your agent can call it, but parsing the output is fragile. Google changes formatting, your parser breaks.
Perplexity Deep Research has clean citations and real-time web access — Perplexity's core strength. But deep research is UI-only. No API endpoint. Your agent literally can't call it.
OpenAI Deep Research requires a $200/month ChatGPT Pro subscription and is also UI-only. The reports are thorough — o3-based reasoning is genuinely good at multi-step research. But there's no API. Your agent has no way to use it.
GPT Researcher and STORM are open-source alternatives you self-host. Full control, no per-query pricing. The tradeoff: self-hosted web crawling is meaningfully worse than what Google or Bing-backed tools can retrieve. Setup is non-trivial. If you have a team that can maintain it and your volume justifies the infrastructure, it's viable. Most teams don't.
What to look for beyond the demo
Consumer deep research demos well because they produce impressive-looking reports. When you're evaluating tools for agent use, the criteria shift:
Can the agent get structured output? Not "can I read the report." Can the agent parse sections, extract citations, and use the findings in the next step of a pipeline? If the tool returns a wall of text, the answer is no.
How dense are the citations? A deep research report without citations linking every claim to a source is just confident hallucination with better formatting. Randomly sample citations on your first few reports. You'll be surprised how often they don't actually support the claim.
Can you control the depth? A quick competitive overview needs 5-10 sources and 2 minutes. A comprehensive landscape analysis needs 50+ sources and 10+ minutes. The tool should let you choose, and tell you the cost before it runs.
Is it a CLI or a UI? This is the filter that eliminates most of the field. If a tool lives in a chat interface, your agent can't use it. End of evaluation.
Where deep research fits in a real workflow
The value of deep research isn't the research itself. It's what happens after.
An agent doing competitive analysis first does deep research on the market landscape. Then it searches for pricing specifics on each competitor it found. Then it generates a comparison infographic. Then it compiles everything into a report and publishes it.
That's four CLI commands, chained together by an agent that understands the goal:
anycap research --query "AI search tools market 2026" --depth comprehensive --output landscape.md
anycap search "competitor-name pricing 2026" --citations --output pricing.json
anycap image generate --prompt "comparison infographic from landscape.md" -o comparison.png
anycap page publish report.md --title "AI Search Tools: Market Analysis 2026"
No SDK. No middleware. Just tools the agent can invoke because they live in its runtime.
Further reading:
- AI-Powered Search for AI Agents: Grounded Search vs RAG — How grounded search and deep research complement each other
- AI Workflow Automation: Build an Agentic Pipeline — Full pipeline: research → generate → publish
- Agentic Analytics Tools in 2026 — How deep research fits into analytics workflows