Multi-Agent Research Pipeline
A reusable scaffold of four specialized LLM agents — Lead Contributor, Positioning, Method, and Experiment owners — that collaborate to draft, review, and assemble a research paper end-to-end.
A workflow we use across our own research projects: each role is an
LLM-backed agent with a fixed set of skills (define structure, narrative
synthesis, write intro, review & merge, consistency check, LaTeX final
assembly, version control). Shared skills like academic search, latex
assembly, and assistant-text extraction live in a sibling shared_skills/
package.
Why agents per role, not per task
A single mega-agent loses coherence across long sections; one agent per role preserves voice, lets us prompt-engineer per-section conventions (notation, citation density, evidence standards), and gives us a clean ownership signal for review and version control.
Where it ships
dcn_v2_agents/— used to produce our DCN-V2 paper.llm_bias_agents/— used to produce our LLM bias paper, with the same four-role structure and additional benchmark/validation skills.
The pipeline is being generalized into a standalone library so future EIL projects can scaffold a paper workflow in one command.