Skip to content
StackCook
Subscribe

Build a deep-research agent

Search the web and sources, read them, and return a cited report.

Early ranking, based on docs and benchmarks, not yet hands-on tested.

Difficulty
Medium
Status
Early
Last reviewed
2026-10-17

Stack

Framework
  1. #1LangGraph7.5Compare
    Why this score for LangGraph(5 notes)
    Fit for the job 8/10
    Supervisor + parallel sub-agent reference builds (Open Deep Research, GPT Researcher).
    Developer experience 6/10
    Lower-level graph API; more code. Editorial read of docs; no own test yet.
    Reliability & production readiness 8/10
    Checkpointing for long runs; ODR ranked #6 on Deep Research Bench (Aug 2025).
    Openness 8/10
    MIT.
    Momentum 7/10
    Core repo active; LangChain archived open_deep_research (last push 2026-08-10).
  2. #2OpenAI Agents SDK6.9Compare
    Why this score for OpenAI Agents SDK(4 notes)
    Fit for the job 7/10
    Official research_bot: planner, parallel search agents, writer.
    Developer experience 7/10
    Editorial read of docs; no own test yet.
    Reliability & production readiness 6/10
    Example repo only; no production deep-research case found.
    Openness 7/10
    MIT; provider portability not verified.
  3. #3smolagents6.7Compare
    Why this score for smolagents(5 notes)
    Fit for the job 7/10
    HF Open Deep Research: CodeAgent reached 55.15% on GAIA validation (Feb 2025).
    Developer experience 7/10
    About 1,000-line core. Editorial read of docs; no own test yet.
    Reliability & production readiness 5/10
    Code actions need a sandbox (Modal/E2B/Docker); no production case found.
    Openness 8/10
    Apache-2.0.
    Momentum 7/10
    Last push 2026-10-06.
Weights
  • Fit for the job: 25%
  • Developer experience: 20%
  • Reliability & production readiness: 20%
  • Openness: 15%
  • Ecosystem & integrations: 10%
  • Momentum: 10%
Model
  1. #1GPT-6.1 Sol7.5Compare
    Why this score for GPT-6.1 Sol(6 notes)
    Task quality 9/10
    95.4% on BrowseComp (Mercor 130); 67.7% on held-out Extended.
    Cost per typical job run 8/10
    $2/$10 per MTok; $0.72 per AA index task.
    Latency / speed 5/10
    66.0 output tokens/s.
    Tool-calling & structured-output reliability 6/10
    70.0% on τ²-Airline.
    Context & modality support 8/10
    1.05M context.
    Openness / deployability 3/10
    Closed weights; OpenAI API.
  2. #2DeepSeek V4.1 Flash7.4Compare
    Why this score for DeepSeek V4.1 Flash(6 notes)
    Task quality 6/10
    85.8% on BrowseComp per an aggregator only; unverified at source.
    Cost per typical job run 9/10
    $0.30/$1.20 per MTok peak (off-peak half); $0.27 per AA index task.
    Latency / speed 8/10
    175.9 output tokens/s.
    Tool-calling & structured-output reliability 8/10
    76.8% on τ²-Airline.
    Context & modality support 8/10
    1M context; image + text input.
    Openness / deployability 8/10
    MIT open weights; self-hosting a 552B MoE needs serious hardware.
  3. #3Claude Sonnet 5.57.2Compare
    Why this score for Claude Sonnet 5.5(6 notes)
    Task quality 8/10
    Top of held-out BrowseComp Extended (73.0%); 86.6% on the original set.
    Cost per typical job run 6/10
    $2/$10 per MTok, but $5.46 per AA index task (verbose).
    Latency / speed 8/10
    141.4 output tokens/s.
    Tool-calling & structured-output reliability 8/10
    78.0% on τ²-Airline.
    Context & modality support 8/10
    1M context.
    Openness / deployability 4/10
    Closed weights; API + Bedrock + Google Cloud.
Weights
  • Task quality: 40% — Report accuracy depends on persistent multi-step search and reasoning.
  • Cost per typical job run: 20%
  • Latency / speed: 5% — Research runs are background jobs; minutes are acceptable.
  • Tool-calling & structured-output reliability: 15%
  • Context & modality support: 10%
  • Openness / deployability: 10%
Search / browse
  1. #1Parallel Search API7.4Compare
    Why this score for Parallel Search API(5 notes)
    Coverage of required actions 8/10
    86.3% on BrowseComp; Search + Extract + Task APIs.
    Security model 7/10
    SOC2 listed for every API; ZDR on enterprise.
    Reliability / maintenance 7/10
    600 req/min on Search.
    Setup effort 8/10
    Free hosted MCP for personal use; $5 free credits/mo.
    Openness / official vs community 6/10
    Official, MIT SDK; proprietary index.
  2. #2Exa7.2Compare
    Why this score for Exa(4 notes)
    Coverage of required actions 7/10
    82.2% on BrowseComp; search + page contents + deep search.
    Security model 7/10
    SOC 2 Type II listed; zero data retention on Enterprise.
    Setup effort 8/10
    Hosted MCP, no API key needed to start.
    Openness / official vs community 7/10
    Official MCP server is MIT open source.
  3. #3Perplexity Search API7.1Compare
    Why this score for Perplexity Search API(4 notes)
    Coverage of required actions 8/10
    Top search config on BrowseComp (89.0%, same model); full-page fetch not verified.
    Security model 6/10
    API key, read-only; SOC 2 / data-retention terms not found.
    Setup effort 8/10
    One REST call; $5 per 1K requests.
    Openness / official vs community 6/10
    Official, Apache-2.0 SDK; proprietary index.
Weights
  • Coverage of required actions: 30%
  • Security model: 25%
  • Reliability / maintenance: 20%
  • Setup effort: 15%
  • Openness / official vs community: 10%
Knowledge
  1. #1Qdrant7.6Compare
    Why this score for Qdrant(5 notes)
    Retrieval quality & features 8/10
    Hybrid and multi-stage queries.
    Operational simplicity 7/10
    Separate service to run.
    Cost at MVP scale 8/10
    Free 1 GB RAM tier forever.
    Scale & performance 7/10
    Vendor claim only.
    Openness 8/10
    Apache-2.0, self-hostable.
  2. #2Supabase (pgvector)7.4Compare
    Why this score for Supabase (pgvector)(4 notes)
    Retrieval quality & features 7/10
    pgvector + hybrid search in Postgres.
    Operational simplicity 8/10
    Same Postgres as the app.
    Cost at MVP scale 8/10
    Free 500 MB; Pro from $25/mo.
    Scale & performance 6/10
    No independent benchmark used.
  3. #3Pinecone7.1Compare
    Why this score for Pinecone(4 notes)
    Operational simplicity 8/10
    Fully managed.
    Cost at MVP scale 7/10
    Free Starter; $20 Builder; $50 min Standard.
    Scale & performance 7/10
    Vendor claim only.
    Openness 3/10
    Proprietary managed service.
Weights
  • Retrieval quality & features: 30% — Deduping and re-finding fetched sources is the main job.
  • Operational simplicity: 20%
  • Cost at MVP scale: 20%
  • Scale & performance: 15%
  • Access control & data governance: 5% — Sources are mostly public web pages; per-user ACLs matter less.
  • Openness: 10%
Evals
  1. #1Langfuse8.0Compare
    Why this score for Langfuse(3 notes)
    Integration effort 8/10
    OTel + framework integrations. Editorial read of docs; no own test yet.
    Cost / free tier 8/10
    Hobby free 50k units/mo.
    Openness / self-host 8/10
    Self-hostable.
  2. #2LangSmith7.1Compare
    Why this score for LangSmith(4 notes)
    Tracing depth & debugging 8/10
    Open Deep Research runs in LangGraph Studio on smith.langchain.com.
    Integration effort 7/10
    Tightest with LangGraph.
    Cost / free tier 7/10
    Free 5k traces; Plus $39/seat.
    Openness / self-host 4/10
    Managed SaaS; self-host enterprise only (not verified).
  3. #3Braintrust6.9Compare
    Why this score for Braintrust(2 notes)
    Cost / free tier 7/10
    Starter $0, 14-day retention; Pro $249/mo.
    Openness / self-host 4/10
    Proprietary.
Weights
  • Tracing depth & debugging: 25%
  • Eval features: 30% — Report quality (citations, coverage) has to be graded with datasets and LLM judges.
  • Integration effort: 15% — All three integrate with the top frameworks; grading matters more.
  • Cost / free tier: 15%
  • Openness / self-host: 15%
Hosting
  1. #1Render7.4Compare
    Why this score for Render(4 notes)
    Fit for workload 8/10
    Background workers for long async jobs; Docker services.
    Cost at MVP scale 7/10
    $7/mo 512 MB instance; free-tier limits for workers not checked.
    Deploy DX 7/10
    Editorial read of docs; no own test yet.
    Lock-in / portability 8/10
    Docker containers portable.
  2. #2Cloudflare Workers6.9Compare
    Why this score for Cloudflare Workers(4 notes)
    Fit for workload 7/10
    Workflows: unlimited wall-clock per step, durable state; CPU up to 5 min per step.
    Cost at MVP scale 8/10
    Free 100k req/day; Paid from $5/mo.
    Deploy DX 6/10
    Step/Workflow model to learn. Editorial read of docs; no own test yet.
    Lock-in / portability 5/10
    Workers runtime differs from Node.
  3. #3Vercel6.5Compare
    Why this score for Vercel(3 notes)
    Fit for workload 6/10
    Hobby caps functions at 300 s; Pro 800 s, 30 min in beta; Workflows for longer.
    Cost at MVP scale 6/10
    Commercial use needs Pro ($20/mo).
    Deploy DX 8/10
    Editorial read of docs; no own test yet.
Weights
  • Fit for workload: 40% — Runs take minutes; long-running background jobs are the core need.
  • Cost at MVP scale: 20%
  • Deploy DX: 15% — Deploy speed matters less than run length here.
  • Reliability / scaling: 15%
  • Lock-in / portability: 10% — Run length matters more than portability for this job.

Get ranking updates for this stack

When a pick changes or a new build lands, it's in the weekly email.

Subscribe