In-Depth Analysis of the Latest Deep Research Technology (huggingface.co)

Since early 2025 major vendors pushed “Deep Research/Deep Search” from experiment to product: OpenAI rolled out Deep Research (Feb) and broadened access with a lightweight tier (Apr), while Google formalized Deep Search inside AI Mode at I/O and stitched it into Gemini 2.5 (rolled into paid tiers from July). The trend is a new search and knowledge-work paradigm where LLM-powered agents perform multi-step reasoning, large-scale networked retrieval, cross-source evidence aggregation and structured, citation-backed reporting — effectively automating literature reviews, hypothesis generation, end-to-end workflows and even agentic actions (e.g., autonomous bookings). Startups, open-source projects and papers multiplied around this capability, signaling a shift in search standards and tooling for research-grade outputs. Technically, architectures have converged on two families: static workflows (human-defined pipelines: ingest → retrieve → parse → summarize) that favor stability and fault-tolerance, and dynamic workflows where agents plan, execute, reflect and adapt. Dynamic systems split into single-agent (end-to-end reasoning with long contexts; ReAct-like examples such as Agent-R1, ReSearch) and multi-agent (planner + specialized executors; OpenManus, deerflow). Core infra mixes search APIs and browser simulation, plus hybrid indexes (keyword + vector). Key engineering challenges are traceable citations and fact verification, evidence selection under conflict, context/memory management for multi-agent coordination, and balancing cost/latency of long inference chains. Practically, most production systems blend static scaffolding with dynamic agent autonomy to trade off reliability, generalization and resource cost.