Deep Research landscape¶

This page surveys the commercial Deep Research products from major AI vendors and the open-source skills/frameworks that try to reproduce similar capability locally. It also explains what this repo vendors (deep-research/) and which adjacent options are deliberately not vendored.

Commercial Deep Research products (2024–2026)¶

The five major vendors converged on a similar shape — multi-agent planning, parallel retrieval, citation-grounded synthesis — but with different defaults around browser scope, model size, and pricing.

Vendor	Product	Launched	Architecture	Notable
OpenAI	ChatGPT Deep Research (`o3-deep-research`, `o4-mini-deep-research`)	2025-02	Plan → multi-step search → cited report	Plus 25/mo, Pro 250/mo; 2026-02 added scoped sites + collaborative planning + live progress
Google	Gemini Deep Research (Interactions API)	2024-12	Lead agent + parallel sub-agents + synthesis agent	Hooks into Gmail/Drive/Chat; 1M-token context; Canvas visualizations; 2026-04 added MCP tool connectors
xAI	Grok DeepSearch / DeeperSearch / Grok 4 Heavy	2024-Q2 → 2025-07	Multi-agent parallel reasoning, live X + web search	256k token context; no formal citations yet; emphasizes raw performance
Anthropic	Claude Research	2025-06 (eng blog)	LeadResearcher + sub-agents + CitationAgent	~90% accuracy uplift over single-agent (Opus 4 lead + Sonnet 4 sub); strict enterprise privacy
Meta	(no first-party DR product)	—	—	LLaMA 4 Scout/Maverick are multimodal foundation models; deep-research must be assembled from LangChain / LlamaIndex / vector DBs

The shared pattern across all four shipping products:

Plan — decompose query into sub-questions
Parallel retrieve — spawn sub-agents that hit web search / docs / files
Iterate — re-plan based on what's found, follow citations
Synthesize — assemble a cited report (markdown / HTML / PDF)

Typical wall time is 2–20 minutes; cost is gated by subscription tier or per-call API metering.

Reproducing Deep Research locally — the six layers¶

A real "deep research" stack for an agent CLI is not a single skill. Decompose it into six layers; pick a tool per layer:

Layer	What it does	Representative tools
1. Planning / decomposition	Turn the question into sub-questions, evidence requirements, and a research plan	`199-biotechnologies/claude-deep-research-skill` (vendored), `langchain-ai/deepagents` (Python framework, not a skill)
2. Retrieval	Web / news / GitHub / PDF search	`tavily-ai/skills/tavily-search`, `firecrawl/cli/skills/firecrawl-search`
3. Extraction	Pull main content, tables, multi-page text out of fetched URLs	`firecrawl-scrape`, `firecrawl-crawl`, `firecrawl-parse`
4. Browser interaction	Login walls, JS rendering, dynamic UIs, downloads	`vercel-labs/agent-browser` (32k⭐), `browser-use/browser-use` (93k⭐)
5. Evidence management	Source registry, claim ledger, citation verification	`199-biotechnologies/claude-deep-research-skill` (built-in)
6. Synthesis	Final report (markdown / HTML / PDF) with inline citations	`199-biotechnologies/claude-deep-research-skill`, `tavily-ai/skills/tavily-research`

The 2-skill minimum that gets you something usable: (1) + (5) + (6) from one disciplined research-flow skill, plus (2) from a search backend you trust.

What this repo vendors¶

Just one skill, deliberately:

vendor/deep-research/deep-research — from 199-biotechnologies/claude-deep-research-skill (646⭐). A pure prompt-flow skill covering planning, evidence management, and synthesis. Four modes: quick (3 phases, 2–5 min), standard (6 phases, 5–10 min, default), deep (8 phases, 10–20 min), ultradeep (8+ phases, 20–45 min). Uses an evidence ledger + claim ledger so each finding traces back to its source.

Why only one? Deep-research stacks built from tavily-ai/skills, firecrawl/cli, vercel-labs/agent-browser, or browser-use/browser-use all introduce paid API keys, hosted backends, or cloud browser sessions. Keeping this repo cost-free by default means anyone can run the vendored research skill against whatever search/browser tools their agent already has (WebSearch, WebFetch, MCP, etc.) without signing up for a third-party service.

If you want a higher-fidelity pipeline, install those skills alongside this one — they compose cleanly.

Adjacent options not vendored — and why¶

Recorded so future agents don't relitigate it:

Option	Stars	Why not vendored
`tavily-ai/skills`	285⭐	Requires Tavily account / API key. Excellent if you're willing to pay; install separately when you are.
`firecrawl/cli`	375⭐	9 skills (search/scrape/crawl/map/extract/agent…). Free tier exists but production use needs a key. Better as a per-project install.
`vercel-labs/agent-browser`	32k⭐	Browser automation is a heavy dependency (Chromium, profiles). Out of scope for a research-only stack; install if you actually need it.
`browser-use/browser-use`	93k⭐	Skills.sh shows a failed Gen Agent Trust Hub security audit. Skip until that resolves.
`langchain-ai/deepagents`	22k⭐	Not a `SKILL.md` repo — it's a Python package / framework. Cannot be vendored as an agent skill. ChatGPT-style summaries sometimes mislabel it.
`24601/agent-deep-research`	4⭐	Genuine Gemini Interactions API wrapper, but too small / experimental to commit to. Worth re-evaluating later.

Things that look like deep-research skills but aren't¶

forrestchang/andrej-karpathy-skills (124k⭐) — name and star count are misleading. It's a single CLAUDE.md distilling four Karpathy observations on LLM coding pitfalls. Not a skill, not about research. See Karpathy's LLM Wiki pattern for the separate knowledge-base pattern Karpathy actually published in his LLM Wiki gist.