DeepAgents

Sweet spot: a product team that wants a "Deep Research"-style feature without building the planner / researcher / synthesiser scaffolding from scratch. DeepAgents packages the pattern competently and lets you focus on tools and prompts — which is where product differentiation actually happens. Reproducing Perplexity Pro's UX in your own app is a one-weekend project with this library plus Tavily. Failure modes. The cost profile is surprising at first — a multi-step research plan with 5 sub-agents and a synthesiser is easily 20+ LLM calls. Without caching and smart model selection (cheap planner, expensive synthesiser), your per-report spend is punitive. Quality also depends on your search tool — Tavily is good, but your real moat is a better (domain-specific) retrieval layer. What to pilot. Generate 10 real research reports on topics where you have expertise. Read them critically. Count citation accuracy, synthesis quality, and hallucinations. If 7+ are genuinely useful, DeepAgents is a strong foundation — now invest in your tools and prompts to hit 9+. If fewer than 5 are useful, the gap is probably model or search; fix those before evaluating the framework again.

DeepAgents is a LangChain-maintained library for building "deep" research agents — agents that decompose a complex question into sub-tasks, research each sub-task in parallel, and synthesise the results into a long-form, citation-backed answer. It is the open-source answer to commercial deep-research products (OpenAI Deep Research, Perplexity Pro). The architecture, built on top of LangGraph, pairs a planner agent that produces a research plan with a research sub-agent spawned per sub-task, each with its own context and tool access. Results are merged and handed to a synthesis agent that writes the final report. Tools are pluggable — web search (Tavily, Exa, Brave), file search, code execution, custom APIs. DeepAgents is under rapid development and ships opinionated defaults: a Tavily-first search stack, GPT-4o-class recommended models, and preset prompts that produce reports in the 1,500-3,000-word range with inline citations. Teams have used it to reproduce a "Perplexity Pro / Deep Research"-style feature in their own products for a fraction of the API cost. MIT-licensed, maintained by the LangChain team, and designed to be extended rather than used as-is. It is a strong starting point for any team building research-assistant features.