What agentic-research does

Writing tooling for papers and proposals, installed once instead of copied

agentic-research is a Claude Code plugin that sets up and polices a paper, proposal or position-paper repository. One command scaffolds the repo, a compiling LaTeX skeleton for the venue, build-on-save scripts, a citation store that checks references are real, and a set of writing agents. It replaced a habit of copying template repositories. The same six agents had ended up in 23 near-identical copies that differed only by a hardcoded path.

Describes v1.8.0 at commit b34bfd5, read on 9 November 2026.

A paper repo's life with the plugin

The six skills

Skill	What it does	Where a human decides
`/research:preflight`	Checks for latexmk, a TeX distribution, the venue's font packages, Node 18+ and pdfinfo. Only git is a hard stop.	Every install is offered one command at a time, with its download size. Nothing installs without a yes.
`/research:establish`	Interviews you, writes the per-repo delta, scaffolds the layout, installs build scripts and CI, seeds the memory bank, and confirms the skeleton compiles. Upgrades an existing repo by diffing, never overwriting.	The interview asks about the research first (topic, core claim, state of the work, audience) and then the paperwork. Any question can be skipped, and the skill may not invent a project brief. It asks before creating a remote or pushing.
`/research:audit`	Reports drift from the pinned version, stale or template-text memory files, citation-matrix drift, and anything that blocks submission.	Report only.
`/research:add-paper`	Fetches one reference by DOI, arXiv id or URL, verifies it against Crossref, arXiv or OpenAlex, and writes the BibTeX and matrix entry.	It refuses to invent an entry it can't verify, which leaves the call to you.
`/research:citation-matrix`	Verifies every cited key: existence, author, title, year and venue against the scholarly APIs; flags preprints, undisclosed self-citations, orphans and missing entries.	Flagged items go to the author.
`/research:originality-check`	Looks for verbatim and close-paraphrase overlap with your sources and cited works, self-plagiarism against your own prior papers, unmarked quotations, and AI-drafted text you haven't verified.	A report for you to judge, never pass or fail. It has no access to published corpora and cannot certify originality.

The eight agents

Agent	Commands
`paper-agent`	outline, survey, draft, revise, create-sprint, validate, camera-ready
`proposal-agent`	analyze-cfp, design, update-design, create-sprint, validate. Nothing is written without an approved design.
`position-paper-agent`	thesis, arguments, survey, draft, stress-test, validate
`latex-agent`	compile, init, check-refs, format-bib, page-count, fix-warnings
`citation-agent`	build, verify, add, check, sync. Owns the citation matrix.
`originality-agent`	review, self-check, unverified
`review-agent`	review, verify-citations, compile-check, update-matrix. The merge-gate report.
`memory-agent`	update, archive, validate, status, sync-phases

Every agent reads its paths from the repository's research delta instead of hardcoding them. That one change is what collapsed the copied variants into one installed set.

What the checker enforces

research-checks.mjs is plain Node and runs the same way for a person, an agent or CI. It fails the build on: a missing or unpinned plugin version, unfilled delta placeholders, placeholder text (TODO, TBD, FIXME) in the paper, a \cite with no bibliography entry, undefined references after compiling, author details left in a double-blind submission, and a delta that disagrees with whether a governance delta is present. It warns on orphan bibliography entries, cited keys missing from the citation matrix, and build output left beside the source.

With agentic-governance

The plugin is standalone. When a repo also runs agentic-governance,establish is supposed to defer to the governance delta for the memory-bank path and print allowlist lines that keep the paper out of the L0 lane. Paper prose is semantic work, so an agent should never be able to merge writing without human review.

Known gaps

Found by reading the source on 9 November 2026. These are the reasons for thecompleteness benchmark.

Composition looks in the wrong place. establish and the checker look for docs/governance-delta.md. Since governance v0.3 that file lives atllm/governance/governance-delta.md. On a current repo, research would conclude governance is absent and skip the deny lines that keep paper prose out of the L0 lane. This is a prediction from the source; the benchmark's first run is set up to confirm or refute it.
Its own delta is in the data plane. docs/research-delta.md governs the repo, so under governance's ADR-0001 it belongs in llm/.
It assumes Claude Code. The skills use Claude-specific tools for asking the user and for locating plugin files, and they are invoked as slash commands. The 1 October audit maps the separation needed for other harnesses. It hasn't been implemented.
It supports writing far more than research. The source never mentions hypotheses, reproducibility, pre-registration, ethics review or responding to reviewers. Experiment and analysis support amounts to optional experiments/, data/ and results/folders. That came from a keyword search, which is a blunt instrument. Measuring it properly is the point of the benchmark.

Sources at commit b34bfd5: README,skills, agents,research-checks.mjs,harness-independence audit.