📖 Overview

Agentic Bootstrap runs a team of AI agents on your dataset, each assigned a different prior-belief persona to simulate virtual researchers. The agents independently answer the same research question, revealing how much a conclusion depends on defensible analytical choices and see the whole agentic garden of forking paths.

You provide a research question and a dataset in CSV format. You get back each agent’s reported finding and every analysis specification it explored along the way, all saved as CSVs for easy comparison.

🚀 Quick Start

⚠️ Prerequisites: Node.js ≥ 18 and Claude authentication — a Claude subscription via ~/.claude, or ANTHROPIC_API_KEY for API-key billing.

⏱️ Runtime & Cost: Each agent run explores ~100 specifications in roughly 15 min. With a Claude subscription, it's covered by your plan. With API billing, it costs about $1.68 per run on claude-sonnet-4-6.

1. Install

git clone https://github.com/jmiao24/AgenticBootstrap.git
cd AgenticBootstrap
npm install
npm link

2. Apply agboot to your dataset

Pass your research question and a CSV to the agboot command line tool. Everything else is automatic.

agboot \
  --question "Does X affect Y?" \
  --data_path /path/to/your/data.csv \
  --persona anti,pro,neutral

--question accepts a literal string or a path to a .md file. By default the agents inspect the data themselves to decide which columns operationalize the exposure and outcome. To give them a curated variable list and effect-size convention instead, write your own data-prompt template and pass it with --data_description (see prompts/data/_generic.md for the format).

Example on immigration data

The repo ships a ready-to-run example for International Social Survey Programme (ISSP) data on immigration and welfare attitudes. From the repo root, run:

agboot \
  --question "How does immigration affect public support for social welfare programs?" \
  --data_path examples/issp/data/data_clean.csv \
  --data_description issp.md \
  --persona anti,pro,neutral \
  --out examples/issp/results

--data_description issp.md uses the paper's exact ISSP prompt; add --dry_run to preview the assembled prompts and personas at no cost. See examples/issp/ for the data and prompt details.

Personas

Personas are one-paragraph prior-belief stances. Pick an opposing pair matched to your debate: pro/anti when people disagree about an effect's direction, or believer/skeptic when they disagree about whether an effect exists. Add neutral for a baseline with no stated prior.

Name	Prior belief
pro	the exposure has a real, beneficial (positive) effect
anti	the exposure has a real, harmful (negative) effect
neutral	no strong prior; let the data speak
believer	the effect is real and substantial
skeptic	the effect is negligible or non-existent

To use fixed personas instead of generating them, put <name>.md files in a directory and pass --persona_path <dir> --persona <names>.

Options

Flag	Default	Description
--question	(required)	Research question: a literal string or a path to a .md file
--data_path	(required)	Path to the dataset (e.g. a CSV)
--persona	anti,pro,neutral	Comma-separated personas (built-in: anti, pro, neutral, believer, skeptic)
--persona_path	(none)	Directory of static <name>.md persona files — skips generation
--data_description	_generic.md	Data-prompt template in prompts/data/ (default lets the agent inspect the data; use issp.md for the ISSP example)
--runs	10	Runs per persona
--rounds	10	Rounds of iterative refinement per run
--model	claude-sonnet-4-6	Model for the agents and persona generation
--concurrency	4	Parallel agent runs
--out	./results	Where outputs are written
--dry_run	(off)	Assemble and print the prompts + personas; run no agents

📁 Output Structure

Each agent run logs all 100 explored specifications and its final selection:

<out>/
├── personas/                           # cached generated persona prompts
├── summary.json                        # per-run status, timing, cost
└── <persona>/run_<id>/
    ├── results_r1.csv ... results_r10.csv   # every spec explored, per round
    ├── results_final.csv                    # the agent's chosen finding
    ├── analysis_final.py                     # reproducible code for the finding
    ├── full_output.txt                       # full agent transcript
    └── metadata.json                         # model, rounds, elapsed, cost

Reading the results (both stdlib-only, no install):

python codes/aggregate.py <out>                 # per-persona mean/median effect, sign split, the gap
python codes/m_value_cal.py <out> --claim -0.35 # m-value of a reported claim vs the agent analysis space

aggregate.py summarizes each persona's final findings; m_value_cal.py pools every logged specification into the analysis-space distribution and reports the 95% Agentic Bootstrap interval and the m-value of a claim (the fraction of defensible analyses at least as extreme).

📚 Citation

@article{miao2026agentic,
  title         = {The Agentic Garden of Forking Paths},
  author        = {Miao, Jiacheng and Pritchard, Jonathan K. and Zou, James},
  year          = {2026},
  eprint        = {2607.01507},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  journal       = {arXiv preprint arXiv:2607.01507}
}

Released under the MIT License.