> ## Documentation Index > Fetch the complete documentation index at: https://labs.prompthon.io/llms.txt > Use this file to discover all available pages before exploring further. # Deep Research Agents

Need help?

Ask in Discord Open a GitHub issue Support guide

## Summary Deep research agents turn an open topic into a managed investigation. Instead of answering from one prompt or one search result, they plan the work, gather evidence in rounds, summarize partial findings, and compile a traceable report. ## Why It Matters Research is one of the clearest product surfaces where agent systems can be useful and dangerous at the same time. Useful, because the work naturally decomposes into search, reading, note-taking and synthesis. Dangerous, because users expect freshness, source quality, and honest handling of uncertainty. A polished answer without a trustworthy evidence trail is not a real research product. ## Mental Model The most durable pattern is a four-part loop: * `plan`: break the topic into bounded research tasks * `collect`: gather evidence for each task * `synthesize`: summarize each task before context overload sets in * `report`: assemble the final answer with explicit sources and open questions This is better thought of as a managed workflow than a single clever prompt. The planner keeps the investigation scoped. The collector expands the evidence set. The summarizer prevents raw search results from flooding the final report. The report writer integrates the results into a reader-facing artifact. ## Architecture Diagram ```mermaid theme={null} flowchart LR Topic["Research topic"] --> Planner["Planner"] Planner --> Tasks["TODO-style task list"] Tasks --> Search["Search and evidence collection"] Search --> Notes["Per-task notes and citations"] Notes --> Summaries["Task summaries"] Summaries --> Report["Final report"] Report --> Review["Human review or delivery"] ``` ## Tool Landscape Deep research systems usually combine a small but opinionated toolset: * search or browsing tools for evidence gathering * notes or artifact storage for structured intermediate state * summarization logic that keeps task outputs bounded * report generation that preserves citations and uncertainty Human checkpoints matter more here than in many other agent products. * planning review prevents wasted search cycles * source review catches weak or irrelevant evidence * final review catches unsupported synthesis and tone drift Artifact strategy also matters. Good systems keep explicit task lists, task summaries, and final reports as inspectable outputs. Those artifacts help with verification, resume flows, and later reuse. ## Tradeoffs * Broad search improves coverage, but it also increases noise and source quality variance. * Aggressive task decomposition improves control, but too many tasks create overhead and repetition. * Strong summarization keeps context clean, but it can hide nuance if citations are not preserved. * A single final answer feels smooth, but a staged artifact trail is more trustworthy. Practical defaults: * plan before searching * keep citations attached at the task level, not only in the final report * treat "unknown" as a valid output * separate evidence gathering from final narrative generation ## Reading Extensions * [Context Engineering](/systems/context-engineering) * [Evaluation And Observability](/systems/evaluation-and-observability) * [Case Studies Overview](/case-studies) ## Update Log * 2026-04-21: Initial repo-native draft based on imported reference material and lab rewrite rules.