1  Introduction: Why Agentic AI Matters for Scholars

1.1 The Goal of This Guide

The goal of this project is simple: to help our accounting academic colleagues and friends get started with agentic AI to aid their research as quickly as possible.

Adopting a new paradigm can be daunting, and it is easy to let the anxiety of complex setups, command-line interfaces, or the “unknowns” of AI lead to procrastination. This guide is designed to:

  1. Dispel Adoption Fear: Show that agentic AI is accessible, manageable, and highly beneficial, dismantling the intimidation barrier.
  2. Demonstrate Real-World Usage: Show exactly how these workflows work in practice, showing how agentic AI acts as a tireless, high-quality research assistant.
  3. Provide a Quick Setup Guide: Offer a clear, zero-friction path to get your own environment running in minutes.

1.2 The Evolution to Agentic Workflows

Chatbots represent the first wave of generative AI, but their utility in serious academic and empirical research is rapidly diminishing. In practice, traditional chat interfaces act as passive conversationalists: they wait for prompts, generate text in isolation, suffer from severe context loss between sessions, and force researchers into an exhausting loop of manual copy-pasting, debugging, and verification.

Agentic AI represents the second stage of this AI evolution. Rather than merely conversing, an agent functions as an active collaborator operating directly inside your project repository. It autonomously reads and modifies files, executes scripts in the terminal, diagnoses errors, and coordinates specialized subagents to enforce rigor and consistency.

Capability What It Means for Academic Research
Read & edit workspace files Makes precise, in-place edits to .tex, .qmd, .R, .py, and .do files
Execute shell commands Compiles LaTeX/Quarto, runs R/Python/Stata scripts, and verifies outputs automatically
Version control integration Creates Git commits, branches, and PRs directly from your instructions
Cross-session persistent memory Retains project rules, notation conventions, and past decisions across sessions (GEMINI.md, MEMORY.md)
Multi-agent review loops Coordinates adversarial worker-critic pairs to catch typos, visual overflows, and econometric errors
Enforced quality gates Validates against explicit quality thresholds (e.g., ≥80 for commit, ≥90 for PR, ≥95 for publication)
Note

For researchers new to developer tools, concepts like Git or command-line scripts might seem unfamiliar. You do not need prior expertise in these tools upfront; a concise, practical introduction to Git and repository concepts will follow.

1.3 The Ladder of AI Coding Tools

To contextualize where agentic workflows fit, it helps to view the ladder of AI tools (Paul Goldsmith-Pinkham, 2026):

  1. Chat Interfaces (Web UIs): ChatGPT, Claude, or Gemini web apps. Great for quick Q&A, but requires manual copy-pasting, lacks local file/repo access, and suffers from context loss between sessions.
  2. Inline Autocomplete / Editors: GitHub Copilot, Cursor, or IDE inline suggestions. Excellent for line-by-line coding assistance, but lacks autonomous multi-step execution, compilation testing, or architectural orchestration.
  3. Agentic Workflows & CLI Harnesses: Google Antigravity or Claude Code. The AI operates directly inside your local repository with file read/edit privileges, terminal execution, subagent orchestration, and persistent project memory.

The transition from Level 1/2 to Level 3 changes AI from a simple line-completer into a full-fledged, autonomous research assistant.

1.4 Why use Google Antigravity as an example?

There is no strong technical reason to use Google Antigravity over others. As summarized in this overview of the agentic coding landscape:

“Six months of convergence has settled the shape of the agentic coding tool and turned the next phase into a contest over the harness, the price, and the habits a team builds around one product.”

This article is a good summary of the current landscape of agentic coding tools. Currently, Claude Code is popular among academics. A key reason we emphasize structured rules, persistent memory, and modular skills is to reduce provider lock-in—making it far easier to adapt your core workflows across tools as platforms evolve.

Note

Healthy competition drives rapid innovation. We ultimately do not know where the future will go or which agentic coding tool will stand the test of time, but the underlying concepts and principles remain universal.

1.5 How It All Works Together

Skills hide most of the mechanics. You describe what you want in plain English, Antigravity figures out which skill fits, and the skill runs the right subagents and checks.

Collaborate via a plan-first approach. Rather than letting the agent write code immediately, establish a concrete plan first. Iterating back and forth with the agent to refine and improve this plan before execution significantly increases the alignment and quality of the final outcome. Once you align on a solid plan, you can step back and let the agent handle the heavy lifting of execution and verification.

1.5.1 What You Do vs What the Skill Does

You Do The Skill Does (once invoked)
Describe what you want Antigravity selects and runs the right skill. Antigravity will come up with the plan
Approve plans The skill runs the orchestrator pattern internally
Review final output Rules load based on files you touch
Say “commit” when ready Subagents fire on the review cycle
TipConcept: What is a “Skill” in Agentic AI?

In an agentic setting, a skill is more than just a prompt; it is a packaged workflow wrapper. While a generic Large Language Model (LLM) knows how to write code, it doesn’t know the specific rules of your project, how to compile Quarto, or how to run your R scripts.

A skill bridges this gap by bundling: 1. Context & Rules: Specific constraints and guidelines for the task (e.g., formatting styles). 2. Orchestration Patterns: Coordinating specialized subagents (like a critic and a fixer) to work together. 3. Tool Access: Executing terminal commands and editing files locally to verify compilation, turning a general-purpose model into a specialized, task-ready executor.

TipConcept: What is a “Commit” and the Safety Layer?

In software development, Git acts as a time machine (version control system) that tracks changes in your files. A commit is a saved snapshot of your project at a specific point in history.

In an agentic workflow, the “commit” phase acts as a critical safety gate. The agent performs edits in your workspace, but these changes are in a temporary, staged state. They are only permanently written to the repository’s official history when you approve them by executing /commit. This human-in-the-loop validation ensures you maintain complete control and that files are only finalized with your explicit approval.

1.5.2 Example: “Convert a working paper into a seminar Beamer deck”

You: "Convert my working paper draft (`Paper/main.tex`) into a 25-minute seminar 
      presentation in LaTeX Beamer. Include the main identification strategy, 
      the baseline regression table, and the mechanism tests."
     ↓
Antigravity invokes econ-slides, which internally:
  → Extracts the research question, empirical tables, and institutional mechanism from `Paper/main.tex`
  → Scaffolds the Beamer deck using institutional preambles (`Preambles/header.tex`)
  → Formats key regression tables and generates clean causal diagrams
  → Runs proofreader (academic tone, notation consistency, math typo checks)
  → Runs slide-auditor / visual-audit (checks for overfull \hbox, line overflow, margins)
  → Compiles XeLaTeX (3-pass build + BibTeX bibliography resolution)
  → Scores the deck against quality gates (target ≥ 90 for conference readiness)
     ↓
You see: "Done. Created `Slides/Seminar_Presentation.pdf` (26 frames, ~25 mins).
          XeLaTeX compiled cleanly with zero overfull hbox warnings.
          Quality Score: 93/100. Ready to review and commit?"

1.6 The Role of the Human in the Era of AI-Assisted Research

As agentic tools lower the friction of coding, running regressions, compiling decks, and drafting prose, what is the human researcher’s role when AI can execute technical tasks autonomously?

1.6.1 Research as Costly Signaling

Academic publishing and promotion operate on costly signaling (Spence, 1973). A signal only separates ability if it is costly to produce. Historically, mechanical execution—writing boilerplate scripts, wrestling with LaTeX syntax, cleaning standard datasets, and drafting routine text—formed a large share of that cost.

When agentic AI drives the marginal cost of mechanical execution toward zero, raw output volume ceases to be a separating signal. The equilibrium inevitably shifts to other, inherently costly human dimensions:

Diminishing Cost Dimension (Automated) Emerging Costly Separating Dimension (Human-Driven)
Boilerplate code syntax and script execution Research Taste & Problem Selection: Formulating novel, economically meaningful questions that matter to theory and practice.
Formatting regression tables and slides Institutional Insight & Domain Nuance: Understanding real-world accounting standards (ASC/IFRS), institutional incentives, and regulatory frictions that AI cannot infer from raw text.
Drafting routine prose and literature summaries Identification Rigor & Causal Logic: Designing airtight identification strategies, stress-testing mechanisms, and anticipating non-obvious confounders.
High-volume surface-level drafts Verification Bandwidth & Rigorous Skepticism: Auditing empirical evidence, stress-testing hypotheses, verifying ground-truth replication, and catching subtle hallucinations.

1.6.2 The Economics of Verification Bandwidth

This dynamic is formalized in recent work on the macroeconomics of AI (Some Simple Economics of AGI, arXiv:2602.20946). The transition to agentic intelligence is characterized by the collision of two racing cost curves:

  1. An exponentially decaying Cost to Automate / Execute: Generating code, models, and text becomes near-free.
  2. A biologically bottlenecked Cost to Verify: Human cognitive attention, critical judgment, and auditing capacity remain strictly finite.

When execution is abundant, human verification bandwidth becomes the binding constraint on scientific progress. Academic value and reputational rents migrate away from pure generation toward underwriting ground truth and assuming liability.

ImportantThe Scholar as the Ultimate Underwriter

An AI agent can execute an analysis, but it cannot bear professional liability. It cannot be held accountable for retracted findings, flawed econometric specifications, or ethical missteps. The human scholar remains the sole legal, moral, and intellectual underwriter of the research.

1.6.3 From Coder to Principal Investigator

Agentic AI does not diminish the human scholar—it elevates you. As emphasized in frameworks like A Causal Affair (Goldsmith-Pinkham, 2026) and Scaling Reproducibility (Xu & Yang, 2026), your role shifts from a line-by-line coder to the Principal Investigator and Chief Architect:

  • You provide the vision, theory, and institutional hypothesis.
  • The agent executes the plan-first implementation, runs diagnostics, and drafts materials within human-defined constraints.
  • You exercise critical skepticism, audit the results against domain reality, and make final decisions.

By adopting agentic workflows, you leverage AI to handle zero-marginal-cost execution, freeing your scarce cognitive bandwidth for the deep, creative, and costly thinking that truly moves the frontier of knowledge.

1.7 Do You Need to Know Everything About Setting Up Agentic Systems?

A common concern among researchers approaching agentic tools is: “Do I need to understand every technical component—all the rules, hooks, and orchestrators—to use agentic AI effectively?”

The short answer is no.

1.7.1 The iPhone Analogy: Helping the Leviathan

When the iPhone first launched, independent developers built countless standalone utilities (flashlights, document scanners, widget dashboards) that Apple eventually absorbed directly into iOS. Early power users and developers were effectively “helping the Leviathan” define modern mobile computing.

A similar convergence is underway in AI. Major frontier labs are packaging scientific capabilities into native platforms—from AlphaFold by Google DeepMind in structural biology to scientific reasoning sandboxes in Claude and ChatGPT.

1.7.2 Why Social Sciences Still Need Bespoke Setups

However, major AI labs focus almost exclusively on the natural and physical sciences, where ground truth is molecular or physical. Empirical accounting and economics—with complex corporate filings, financial databases (WRDS/CRSP/Compustat), institutional accounting standards (ASC/IFRS), and journal replication rules—have not received turnkey solutions. This template fills that gap, providing researchers with the specialized scaffolding that generalist models omit.

1.7.3 Customization as a Durable Skill

Even as future AI models absorb more workflow features natively, knowing how to customize your setup remains an essential skill.

Just as an iPhone user doesn’t need to write iOS kernel code to configure settings, permissions, and shortcuts, an academic researcher doesn’t need to build AI tools from scratch. You only need to know how to calibrate the dashboard—setting project conventions in GEMINI.md, providing institutional domain context, and enforcing the quality gates that match your discipline’s standards.