Multi-Agent Academic Paper Workflow

A multi-agent workflow for writing and iteratively reviewing an academic economics paper, using independent AI instances for writing, quality gates, and simulated peer review.

Paper

"Creative Destruction in the Market for Intelligence: Demand Reallocation When New LLMs Enter"

Using daily data on 385 models from 66 firms over 93 days on OpenRouter, we estimate a nested logit demand model and conduct event studies around model entry events. The headline finding: creative destruction operates almost exclusively through within-family upgrades — predecessors lose 24–35% of daily requests when successors launch, while cross-firm entries produce no detectable displacement.

What This Repo Contains

Directory	Contents
WORKFLOW_v2.md	Paper writing workflow — 5-layer quality assurance with 4 independent critic gates
WORKFLOW_REVIEW.md	Simulated peer review workflow — 5 independent referee roles × 3 rounds
paper/	LaTeX source and PDFs for all versions (original, R1 revision, R2 revision, clean final)
code/	Python analysis scripts (data exploration, construction, regressions, robustness, heterogeneity, number verification)
output/	Regression tables (CSV), figures (PNG), identification memo, pre-analysis plan
logs/	All referee reports, editorial letters, response letters, score tracking, revision changelog
logs/review/	Complete Round 1–3 referee reports, editorial decisions, and response letters
config/	Style references and research rules

Workflow Design

Phase 1: Paper Writing (WORKFLOW_v2)

1. Data Exploration → descriptive statistics, visualizations
2. Research Design → identification strategy, pre-analysis plan
3. Empirical Analysis → regressions, robustness checks
4. Paper Writing → section-by-section LaTeX generation
5. 4 Independent Critic Gates → each section must score ≥7/10

Phase 2: Simulated Peer Review (WORKFLOW_REVIEW)

1. 5 Independent Referees: Field Expert, Methodologist, Writing Specialist, Policy Expert, Red Team (adversarial)
2. Editor synthesizes all reports into an editorial letter
3. Response Letter written to address each point
4. Revisions implemented with version tracking
5. Up to 3 rounds until consensus Accept

Key Results

• Round 1: Average score 6.3/10 → Major Revision
• Round 2: Average score 7.3/10 → Minor Revision
• Round 3: Average score 7.4/10 → Accept
• Total referee reports generated: 13
• Key improvements: identification discussion, robustness (HonestDiD, Sun & Abraham), title and framing

Note

Raw data (data/) is not included in this repository. The code and analysis scripts reference data files that would need to be obtained separately from OpenRouter's public API.