Agentic AI and Document Synthesis

Note: This was Pete Kaminski's main check-in topic during the OGM 2025-11-06 call. This page provides context on what agentic AI is and its applications.

What is Agentic AI?

Agentic AI refers to AI systems that can:

Act autonomously toward goals
Break down complex tasks into subtasks
Create and manage their own to-do lists
Read and write multiple files
Work through problems systematically

Pete describes it as working "about at the level of a human software developer."

Key Difference from Simple LLMs

Simple LLM Chat:

Single request → single response
No task breakdown
No file management
No persistent workflow

Agentic AI:

Can be told: "Fix this bug" or "Build this feature"
Creates own task list
Works through tasks one by one
Checks items off
Reads/writes many files as needed
More autonomous and self-directed

Tools Discussed

Claude Code

Made by Anthropic
Terminal version (for developers)
Web version (more accessible, works on mobile)
Pete has been using for 6-8+ months
$250 credit promotion mentioned (expires Nov 18, 2025)

Droid

Made by Factory Eye
Also has web version
Another major player in the space

Applications

Software Development (Pete's primary use)

Complex, multi-file codebases
Bug fixes
Feature implementation
Refactoring

Document Synthesis

Pete's emerging interest - applying same approach to:

Legal Documents:

Large document sets
Documents that need to fit together
Analysis across multiple contracts
Rewriting with consistency

Research Studies:

Analyzing large corpora
Synthesizing findings
Creating overviews

Serial Fiction:

Understanding large story arcs
Writing new episodes consistent with canon
Planning new volumes in a series
Maintaining character consistency

Framework Synthesis (see Field of Sheaves):

Smashing together related frameworks
Creating 10-20 page overviews
Connecting ideas across domains

Pete's Vision

"Everybody's gonna be doing this kind of stuff in 18 months, 24 months."

Currently: "You could be doing it today" but with challenges.

Challenges to Adoption

For Non-Technical Users

Pete's "Oven Mitts" Observation: Watching someone use these tools, "it was kind of like they had oven mitts on, and they were trying to do fine-grained tasks"

Why It's Hard:

Built by software developers for software developers
High initial overhead
Feels like "molasses" even with easier web versions
Still requires some technical comfort

The Gap

"Not quite ready for mass adoption"
"Easier on-ramp" but still challenging
Web version paradoxically harder than terminal in some ways

Pete's Offer

"I would love to show some people how to do that more."

Seeking collaborators interested in exploring these tools for non-code applications.

The "Field of Sheaves" Project

Pete's specific application: Creating a "meta-collection" of smashed-together frameworks called Field of Sheaves.

Examples created:

Collective efficacy
Social cyclic theories
Intention-action gap
Strategy and management frameworks
Government governance principles
Commons stewardship by communities

Status: Between Amazing and Embarrassing

Pete's ambivalence:

"Half amazing" - successfully synthesizing complex knowledge
"Half embarrassing" - "mostly built by AI"
Too embarrassed to release publicly yet
Working on framing what's good vs problematic

Broader Implications

For Knowledge Work

Massive increase in synthesis capacity
AI as thought partner, not just tool
Human becomes curator/director

For Expertise

Changes what expertise means
Shifts toward orchestration and judgment
Questions of attribution and authorship

For Collaboration

New ways to work with large bodies of text
Team members can leverage huge resources
Democratization of certain types of analysis

Site Navigation