Deep Research Agent: a literature review as a gated, seven-phase assembly line

A review pipeline with gates, not vibes

The project's own summary line is accurate: an autonomous pipeline for rigorous academic research with strict DOI verification, multi-agent Scopus and OpenAlex and Semantic Scholar retrieval, and APA 7th Word document generation. It is written as a skill for the Hermes framework or a compatible ECC or Claude Code runner, and it covers the whole arc from query to compiled document to knowledge-base ingestion.

What distinguishes it from a prompted chat session is that the process is hard-coded in phases, and the documentation states that skipping phases or taking shortcuts, like drafting without full-text verification, is strictly forbidden. Whether an agent reliably honours such prohibitions is an empirical question, but encoding the order of operations into the skill is the right place for rigour to live.

The flow begins with a plan the agent formulates and then halts for your explicit approval, which is the correct first gate: a literature search you cannot steer is a literature search you cannot trust. After approval, a background process scrapes the full text of exactly thirty high-impact papers via APIs while the visible phases proceed.

Discovery, and a blunt quality gate

Phase 1 deploys exactly three concurrent subagents querying Scopus, Exa, OpenAlex and Semantic Scholar for exhaustive mapping. The two-stage sourcing model sits on top of that: subagents first fetch and screen a large pool of abstracts, a hundred or more, and for the highly relevant subset, the agent downloads and deeply reads the full text, specifically the methodology and results sections, to prevent abstract-induced hallucination. The documentation calls full-text verification of final claims absolutely mandatory, and repeats it.

The journal policy is aggressive and worth quoting as policy: explicit targeting of Q1 and Q2 journals, Q3 marked only when strictly necessary, and a blanket ban on Q4 and all MDPI publications. As a review comment, a wholesale MDPI ban is a blunt instrument that removes some legitimate venues along with the volume mills, and any user should decide whether that prior matches their field before inheriting it silently.

The anti-hallucination machinery

The verification phases are the product's spine. Phase 4 strips AI vocabulary from the draft. Phase 4.5 runs live HTTP tests against every generated DOI, and a DOI that does not resolve, a 404, deletes the citation. Generated claims are cross-referenced against the raw full texts to prevent overstatement, which is the check most homegrown literature workflows never build.

Phase 5, named the Remi review after the author's academic advisor, is an internal peer-review loop that critiques and rewrites the draft until academic standards are met, with the stated mission of aggressively removing AI filler phrasing; the README names the same two notorious filler words that style guides love to hate. Phase 6 compiles the final Word document, with Python drawing Mermaid and Matplotlib charts, and is only allowed to run after the verification and review phases have passed. Phase 7 writes research summaries into an Obsidian vault and uploads every individually cited reference as a separate source into Google NotebookLM.

The output contract is strict in one more way: all academic reports and drafts are generated in English regardless of the conversational language, which removes a whole class of mixed-language citation formatting bugs.

Installation and wiring

The skill clones into an agent's skills directory:

cd <AGENT_SKILLS_DIR>
git clone https://github.com/CYC2002tommy/Deep-Research-Agent.git

Python dependencies for document generation and link verification:

pip install python-docx PyMuPDF requests matplotlib seaborn pandas

The host needs Node 18 or newer with npx for MCP servers and Python 3.10 or newer. The MCP layer is where the dependencies bite: Scopus retrieval requires the scopus-mcp server with an Elsevier API key from the developer portal; Phase 7 requires the NotebookLM MCP server with a one-time npx authentication against your Google account; Exa search is recommended as the neural fallback; GitHub and Playwright MCP servers are optional extras.

Two configuration details deserve attention before the first run. The skill looks for an Obsidian vault under a Hermes subfolder with a fallback path under the user profile, overridable through an environment variable or a skill edit. And the documented default output location is a hardcoded drive path from the author's own machine, with configuration offered as the escape hatch; on any other computer, that setting needs changing before anything else.

The paywall problem, stated plainly

The prerequisite the documentation marks as important is also the one that needs the most judgment: Phase 4.6, Zotero archiving and full-text PDF download, uses a browser automation layer named cloakbrowser to get past Cloudflare and publisher paywalls at outlets like Wiley and Elsevier, and the README recommends connecting through a university network for native IP-based authorization, which it says provides a 100 percent success rate for automated PDF extraction.

Both halves of that sentence should be read carefully. Institutional network access is the legitimate route, and using an agent inside it to fetch papers your institution licenses is defensible. The bundled bypass of Cloudflare and paywalls is another matter: at many institutions it violates acceptable-use policy, and against publishers it violates terms of service. Teams adopting this skill should decide where their line is before the agent reaches Phase 4.6, not after.

The project's own ethics disclaimer points the same direction: it is an AI assistant, not an author, and its principles start from human accountability for the work produced. Anti-hallucination machinery reduces one class of error; it does not move the responsibility.

Boundaries, licence, and the manual route

The boundaries are the dependencies. Without a Scopus key the premium retrieval tier is gone; without NotebookLM authentication Phase 7 degrades; without an academic network the full-text pipeline leans on the bypass above. The skill is English-output-only by design, targets the Hermes and ECC ecosystem rather than being a standalone app, and has no published releases, so pinning means pinning a commit. The licence is MIT and the last push was on 2026-08-23.

The manual alternative is the thing it automates: Zotero for archiving, Scopus and OpenAlex searches by hand, a reference manager, and your own reading, with weeks instead of hours. Hosted research assistants offer a middle ground without the phase gates. What this project uniquely sells is enforced process: the citation that would have slipped through gets deleted by a live DOI check, and the draft that would have shipped fluffy gets rewritten by a review loop before compilation is even permitted. If your literature reviews fail on process rather than on knowledge, that enforcement is the feature; if they fail on knowledge, no pipeline fixes that.

Editorial conclusion

Deep Research Agent fits researchers already running Hermes, ECC or Claude Code tooling who want literature reviews produced under enforced phase gates, mandatory full-text verification and live DOI checks, and who have legitimate access to the corpora it queries. It does not fit anyone without Scopus and NotebookLM credentials, anyone who needs non-English output, or teams unwilling to decide in advance where they stand on its paywall-bypass dependency. Verify first: that the Scopus key and NotebookLM authentication work in your environment, that the hardcoded output path is overridden on your machine, and that Phase 4.6 operates inside access your institution actually licenses. The licence is MIT, and the last push was on 2026-08-23.

Frequently asked questions

Official sources

1. CYC2002tommy/Deep-Research-Agent on GitHub
2. Issues
3. License: MIT
4. README