Brian Jagger
I have been building AI research agents for nearly two years, and the shift in what they can do over that window has been genuinely startling. The first generation of agents I built could string together a few web searches and produce a passable summary if you were generous about citation quality. The current generation runs sustained multi hour investigations, synthesises across hundreds of sources, identifies its own gaps, and produces reports that compare favourably with junior analyst work. The category has its own name now, deep research, and the major model providers have all shipped products in it.
AI Research Agents: Building a Deep Research Tool
In this article I want to do two things. First, I want to explain how the leading deep research products actually work behind the curtain. ChatGPT Deep Research, Perplexity Deep Research, and Claude’s research mode each take meaningfully different architectural approaches, and the trade offs are worth understanding before you decide where to invest. Second, I want to give you a concrete blueprint for building your own AI research agent using LangGraph and standard web tools. The pattern is more accessible than people assume, and the customisation you get from owning the loop is meaningful when you have specific domain requirements. I will also share my own evaluation framework, because the most underrated skill in this space is judging research quality with discipline.
Why deep research is the breakout agentic use case of 2026
Research is the cleanest demonstration of what an agentic loop can do. The work involves planning, searching, reading, reasoning, synthesising, and iterating, all of which benefit from being able to use tools dynamically. A static prompt to a chat model produces a confident sounding answer that may be entirely fabricated. An agentic loop with web access produces an answer grounded in actual sources, with citations a human can verify.
The market response has been swift. ChatGPT Deep Research, launched in early 2025, hit ten million users within months. Perplexity expanded into deep research as the second pillar of its product. Claude added research mode and made it free for Pro users. Anthropic’s research stack is integrated into Claude.ai with the ability to read files, browse the web, and reason across sources for tens of minutes per task.
The reason this category exploded is that knowledge workers spend a quarter of their time on research, and the agents are now good enough to compress hours of work into minutes without sacrificing rigour.
The economic case is straightforward. A research analyst at a consulting firm bills five hundred dollars an hour. A deep research run on ChatGPT or Claude costs cents. Even at twenty percent of analyst quality the agent wins on speed. At seventy percent quality, which is roughly where the leaders are in 2026, the agent wins decisively for most use cases.
The canonical deep research loop
Every deep research agent I have studied implements some variant of the same loop. The loop has six stages. Plan, search, read, synthesise, critique, expand. The agent runs through this loop multiple times before producing a final report.
In the plan stage the agent decomposes the question into sub questions. A question like what is the state of solid state batteries in 2026 might decompose into manufacturing readiness, leading companies, technical challenges, recent funding, and projected commercialisation timelines.
In the search stage the agent runs queries across web search, academic databases, and any internal sources it has access to. For each sub question the agent typically issues three to ten queries with different phrasings to maximise recall.
In the read stage the agent fetches the most promising sources, parses them, and extracts the relevant claims. This is where retrieval quality matters most because not every search result is worth a full read.
In the synthesise stage the agent integrates findings into a coherent narrative organised by sub question. In the critique stage the agent checks the synthesis for gaps, contradictions, and unsupported claims. In the expand stage the agent identifies what is missing and decides whether to launch another loop.
| Stage | Output | Failure mode |
|---|---|---|
| Plan | Sub questions | Shallow decomposition |
| Search | Candidate sources | Bad query formulation |
| Read | Claims with citations | Skipping low ranked but high quality sources |
| Synthesise | Coherent narrative | Hallucinated bridging |
| Critique | Gap and contradiction list | Insufficient self skepticism |
| Expand | New search directions | Premature termination |
ChatGPT Deep Research under the hood
ChatGPT Deep Research uses a fine tuned version of OpenAI’s reasoning model series with extended tool use. The product was trained on research workflows specifically, which is part of why its output reads as more report like than a generic chat answer.
The architecture exposes a web browsing tool, a Python execution environment, and the ability to read uploaded files. The agent’s planning is visible to the user through the thinking traces that stream as the research progresses. You can watch it pivot when it discovers something interesting or refine a query when initial results are weak.
What ChatGPT Deep Research does well is breadth. It will fan out to dozens of sources, including academic papers, news articles, government reports, and analyst notes. The synthesis is generally strong, with a clear structure and reasonable handling of conflicting information.
Where it shows weakness is in citation precision. The product cites generously but the link between specific claims and specific sources can be loose. I have audited reports where the cited source did not actually support the claim, or where the source was outdated by several years. For high stakes use cases I always verify citations manually.
The latency is the other constraint. A typical research run takes five to thirty minutes. That is fast compared to a human analyst but slow enough that you batch the work rather than running interactively.
Perplexity Deep Research and the sonar architecture
Perplexity took a different approach. The company started as a search first product and has retained that orientation. Deep Research on Perplexity is built around their sonar models, which are tuned for grounded retrieval and citation density.
The architectural difference shows in the output. Perplexity reports are denser with citations, often citing four or five sources per paragraph rather than one or two. The trade off is that the synthesis can read more like a literature review than a narrative report.
Where Perplexity excels is in current events and rapidly changing topics. The platform’s search index seems particularly fresh, and the agent surfaces information from the last twenty four hours more reliably than its peers. For financial markets, news cycles, and emerging technology I default to Perplexity.
The user experience emphasises the source list throughout the interaction. You can see the agent’s working sources as it researches, and the final report sits next to a curated bibliography you can browse. That transparency builds trust quickly with users who want to verify rather than accept.
Claude research mode and the citation philosophy
Claude’s research mode launched later than the others but with a distinct philosophical approach. Anthropic emphasised citation reliability and epistemic humility over raw capability claims. The product is more conservative about making strong assertions when the evidence is thin.
The architecture combines extended context windows, agentic browsing, and a citation system that ties claims to source spans rather than just source URLs. When Claude cites, you can see the exact passage that supports the claim, not just the page it came from. That precision is rare in the category.
Where Claude research shines is in long form analytical work and tasks that require careful reasoning across many documents. The model handles uploaded research dossiers well, integrating personal documents with web sources in the same investigation.
The honest critique is that Claude research can be slower and more conservative than competitors on tasks where speed and breadth matter most. The same epistemic humility that makes the citations trustworthy also makes the agent more likely to flag uncertainty than to offer a confident synthesis.
I tell teams to think about which Claude virtues match their use case. For regulatory or scientific work the precision is invaluable. For brainstorming or strategy exploration the conservative tone can frustrate.
Building your own with LangGraph and web tools
The major products are excellent for general purpose research. For domain specialisation, you often want to build your own. The framework I use most often in 2026 is LangGraph, paired with web tools from providers like Tavily, Serper, or Exa.
The graph structure of LangGraph maps cleanly onto the research loop. Each stage becomes a node. The transitions between nodes carry the shared state, which includes the original question, the running list of sub questions, the gathered sources, the working synthesis, and the critique log.
A minimal implementation has a planner node, a searcher node, a reader node, a synthesiser node, a critic node, and a router that decides whether to loop again or finalise. Each node is a focused LLM call with structured outputs. The total code base is usually under a thousand lines.
The customisation you gain matters when you have domain specific source preferences. A legal research agent should prioritise statutory sources and case law databases. A medical research agent should weight peer reviewed literature heavily. A competitive intelligence agent should rank investor relations content and 10K filings. Owning the loop lets you encode these preferences explicitly.
- LangGraph for orchestration
- Tavily or Exa for high quality search
- Playwright or Browserbase for full page fetching
- Trafilatura or similar for clean content extraction
- A reranker for source prioritisation
Search, parse, synthesise, critique, expand
The implementation details matter. Let me walk through each stage with the patterns I have found most reliable.
Search benefits from query diversity. I generate three to five queries per sub question, varying phrasing, specificity, and source type. The agent then deduplicates the results before reading. This costs a little more but lifts recall meaningfully.
Parsing benefits from extraction templates. Rather than feeding raw HTML to the model I extract the main content, strip boilerplate, and pass a clean text representation. Tools like Trafilatura or Mercury do this well.
Synthesis benefits from structured intermediate representations. Instead of asking the model to write the report directly I have it produce a structured outline first, with each section linked to specific source claims. The final report is generated from that outline.
Critique is the stage most teams skip and the one I find most valuable. The critique node reviews the synthesis with a critical eye, looking for gaps, contradictions, unsupported claims, and weak reasoning. The output is a structured list of issues that feed the expand stage.
Expand is where the loop closes. If the critique identifies serious gaps and the budget allows, the agent runs another search and synthesis cycle. If not, it cleans up the report and returns.
Source ranking and trust calibration
Not all sources deserve equal weight. The research agents that produce trustworthy output have explicit source ranking logic.
The dimensions I score sources on include authority, currency, primary versus secondary, and corroboration. Authority captures whether the source is a recognised expert, a major news organisation, an official government body, or a random blog. Currency captures recency, which matters more for fast moving topics. Primary versus secondary distinguishes original research from commentary. Corroboration captures whether claims appear consistently across multiple sources.
| Source type | Default authority weight | Notes |
|---|---|---|
| Peer reviewed paper | High | Adjust for journal quality |
| Government report | High | Adjust for political bias |
| Major news organisation | Medium high | Adjust for beat reliability |
| Industry analyst report | Medium | Vendor influence to consider |
| Corporate blog | Medium low | Marketing motivation |
| Personal blog | Low | Case by case |
The weights are encoded as part of the system prompt or as a separate scoring step. The agent uses them when synthesising, preferring high authority sources for substantive claims and using lower authority sources for context or colour.
Citation patterns that hold up under scrutiny
A research report without verifiable citations is worth less than one with them, full stop. The citation patterns I require are precise linking, balanced sourcing, and audit trail.
Precise linking means each substantive claim cites a specific source, ideally with a passage reference. Inline numeric citations linked to a bibliography is the standard. Bracketed footnotes work too. The pattern to avoid is the general bibliography at the end with no inline references, which makes verification impossible.
Balanced sourcing means controversial or contested claims cite multiple sources representing different perspectives. A research agent that cites only one side of a debate is not doing research, it is doing advocacy.
Audit trail means the agent’s process is recoverable. Which searches were run, which sources were fetched, which were used in synthesis, and which were excluded. The audit trail lets you debug bad reports and build trust in good ones.
The single highest leverage improvement most teams can make to their research agent is forcing strict inline citation, with the agent unable to produce a final report that has uncited substantive claims.
Evaluation: citation quality, factuality, novelty
Evaluation is the unsexy work that separates serious teams from demoware teams. The evaluation framework I use has three dimensions.
Citation quality measures whether the citations are precise, accurate, and authoritative. I sample a percentage of citations and check whether the cited source actually supports the claim. I track the rate of broken or irrelevant citations.
Factuality measures whether the claims in the report are true. This requires harder verification work, often involving subject matter experts or comparison to authoritative ground truth sources. I sample claims rather than checking everything.
Novelty measures whether the report adds value beyond what a person could find with a single web search. A report that just summarises the first page of Google results is not deep research. A report that integrates across domains and surfaces non obvious connections is.
LLM as judge can handle the first two dimensions reasonably well in 2026, particularly with high quality reference judges like Claude or GPT 5 doing the grading. Novelty is harder to automate and usually requires human review.
Domain specialisation for legal, medical, and financial
Generic research agents struggle in domains with specialised source ecosystems. The investment in domain specialisation pays back quickly for teams operating in legal, medical, or financial contexts.
For legal research the specialisation includes preferred databases like Westlaw, Lexis, and PACER, citation formats like Bluebook, and jurisdiction aware ranking. The agent should understand the hierarchy of legal authority and weight a Supreme Court ruling above a district court decision.
For medical research the specialisation includes PubMed access, Cochrane reviews, MeSH terms, and evidence ranking that distinguishes randomised controlled trials from observational studies. The agent should be cautious about clinical claims and explicit about evidence strength.
For financial research the specialisation includes 10K and 10Q filings, earnings transcripts, analyst notes, and market data feeds. The agent should be able to compute basic financial metrics and verify them against the filings rather than reproducing analyst commentary uncritically.
Each domain also has compliance considerations. Legal research must respect attorney work product privilege. Medical research must handle PHI carefully. Financial research must avoid generating material that crosses into investment advice without appropriate disclaimers.
Cost, latency, and the patience budget
Deep research is expensive compared to a chat completion. A thorough run can consume hundreds of thousands of tokens, hit dozens of web pages, and take twenty minutes of wall clock time.
The cost economics depend on the model used, the number of loop iterations, and the volume of source content processed. A single research run on a major platform typically costs between one and ten dollars in compute. That is cheap compared to human research but it is not free.
The patience budget is the wall clock time users will tolerate. For interactive use cases under sixty seconds, you cannot do deep research, you need a faster but shallower mode. For asynchronous use cases up to thirty minutes, deep research is viable. For overnight analysis up to several hours, you can run multi loop agents that interrogate the problem from many angles.
I configure my agents with explicit budget constraints. Token budget, time budget, and loop iteration limit. The agent monitors usage and finalises gracefully when limits approach. The graceful finalisation pattern is critical because an agent that runs over budget produces worse output than one that finalised promptly.
Common failure modes in research agents
I have catalogued the failures I see most often. The top patterns are surface skimming, citation drift, perspective collapse, terminator avoidance, and stale information leakage.
Surface skimming is when the agent reads only the first paragraph or the abstract of each source rather than the substantive content. The fix is to chunk the source and have the agent read several chunks before extracting claims.
Citation drift is when the agent’s final report cites sources that contributed little to the synthesis. The fix is to track which sources support which claims at the synthesis stage and prune the bibliography accordingly.
Perspective collapse is when the agent settles on one view and stops searching for counterarguments. The fix is an explicit step in the critique node that demands the agent search for opposing perspectives before finalising.
Terminator avoidance is when the agent keeps looping past the point of useful return. The fix is loop budget and convergence criteria.
Stale information leakage is when the agent fills knowledge gaps with model priors rather than current sources. The fix is grounding constraints that require citations for any substantive claim.
Comparing the major products head to head
Let me give you my honest head to head as of mid 2026.
ChatGPT Deep Research is the strongest general purpose product. The breadth, the structure, and the writing quality lead the field. Best default for general business research.
Perplexity Deep Research is the best for current events, news cycles, and rapidly changing topics. The citation density and source transparency are valuable for verification heavy work.
Claude research mode is the best for high stakes analytical work where citation precision and epistemic humility matter. The integration with uploaded documents is particularly strong.
Google’s Deep Research, on Gemini, has improved meaningfully in 2026 and is the best for tasks that benefit from Google’s search index and integration with Google Workspace. Worth a serious look for users embedded in the Google ecosystem.
For domain specialised work, custom builds on LangGraph with appropriate source connectors will outperform any generic product. The choice depends on whether your domain requirements justify the engineering investment.
Where research agents are going next
The frontier I am watching closely is multi modal research, where agents process not just text but tables, charts, PDFs with figures, and even video transcripts. Early results suggest meaningful capability gains when the agent can read a figure rather than parsing only the figure caption.
Another frontier is multi agent collaboration on research tasks. The pattern of one orchestrator dispatching multiple researcher agents to investigate sub questions in parallel produces better results than sequential single agent work. The orchestration complexity is real but the leverage is significant.
Long horizon research is the third frontier. Today’s deep research runs take minutes to hours. The next generation will run for days, returning to the problem repeatedly as new sources emerge, building up an evolving understanding rather than producing a single static report.
The combination of these directions points toward research agents that function more like ongoing research assistants than one shot tools. The implications for analyst work, journalism, scientific research, and intelligence work are substantial.
