The Collective

Agentic
Frameworks Lab

Agentic Frameworks Lab is a collective of doctoral researchers and advanced practitioners working at the intersection of AI systems, human-computer interaction, and applied machine learning. Our members are completing PhDs across institutions in North America and Europe.

The collective was formed around a shared frustration: the available literature on practical AI use was not written by people who also study the underlying systems. We sit in both positions. We read the research, attend the technical workshops, and have colleagues who build these models. We also use these tools daily for real research work. That dual vantage point is what this guide reflects.

From the Preface

Written.
Not generated.

This book was written by researchers. Not generated, not assembled from machine output, not lightly edited AI text dressed up in a human voice. Written — by the doctoral students and practitioners whose research areas are listed on this page, drawing on their expertise, their direct experience with the systems described, and their access to the research literature that underlies them.

We say this directly because the question will occur to anyone who picks up a book about AI tools: was this written by the thing it's about? The answer is no. The writing, the structure, the judgments about what matters and what doesn't, the decisions about where to hedge and where to be definitive — those are human.

That said, there is one place where Claude was used in producing this book, and we want to be clear about it: testing. Every prompt template was run. Every code example was executed. Every claimed behavior was verified in a live session. Every feature described was checked against current documentation. Where we say something works, it was tested and confirmed to work. Where we note limitations, those limitations were observed directly.

A book about cooking should be written by someone who cooks. A book about a power tool should be written by someone who has used it on real projects. A book about Claude should be written by people who have run thousands of sessions, hit the failure modes, developed the intuitions, and can distinguish what works in practice from what merely sounds plausible. You cannot do that without using the system. You cannot write honest practical guidance about a technology you have only read about.

What We Committed To

Three standards.
No exceptions.

01
Accuracy Over Confidence

The AI space has a surplus of enthusiastic description and a shortage of honest assessment. We tried to write the guide we'd want from a knowledgeable colleague who would tell us what doesn't work as advertised, what's a research preview versus a finished feature, and what the research actually supports. That means acknowledging hallucination as a real and persistent problem. It means saying "we don't know" when we don't.

02
Practical Depth

Every technique in this book should be usable by someone who reads the chapter and tries it the same day. We cut anything that was interesting in principle but vague in practice. The prompt templates were developed and used by our members on real work. The code examples run. The workflow descriptions were tested against actual sessions, not reconstructed from memory or imagined from documentation.

03
Appropriate Calibration

Claude is a rapidly evolving system. Features that are research previews today may be production-stable or significantly changed by the time you read this. We structured the book so that the conceptual content — how to think about prompting, how to approach agentic workflows — is durable, while being explicit about which specific details are likely to change and where to check for current information.

Research Areas

Eight research tracks.
Each chapter reviewed by experts.

Research Area
Focus
Primary Contributions
Agentic AI & human oversight
Autonomous system behavior, supervision protocols
Advanced & Agentic section · Claude Code · Permission modes
Prompt engineering & LLM evaluation
Formal methods for prompt design, output assessment
Foundations · Prompting chapters · Template appendices
HCI & AI-assisted knowledge work
Human-AI collaboration patterns, productivity research
Core skills · Cowork · Productivity applications
NLP & language model behavior
Model internals, emergent behaviors, failure modes
Context windows · Ethics & limits · Responsible use
Multi-agent systems
Orchestration, coordination, emergent system behaviors
Agent teams · Mixed-model configurations · Agentic workflows
AI safety & applied ethics
Deployment risks, governance frameworks, responsible AI
Limits & responsibility · Anti-patterns · Future of field
Applied ML & developer tooling
Production ML systems, developer experience, tooling
Coding · Developer playbook · API reference · Vibe coding
Information retrieval & search
RAG systems, search integration, knowledge grounding
Web search · Research & analysis · MCP integration
Editorial Standards

How we ensured quality

Testing
Every example verified

Every prompt template was run in a live Claude session. Every code example was executed. Every claimed behavior was verified against current Claude behavior — not inferred from documentation or reconstructed from memory.

Review
Multi-author review

No chapter is the work of a single author. Every chapter was reviewed by at least two collective members with relevant expertise. Technical chapters received technical review. Applied chapters received practitioner review.

Calibration
Honest about uncertainty

When we say Claude "understands" something, or "reasons" through a problem, we are describing observed behavior — not making philosophical claims about inner states. The hard questions are open. We don't pretend they're settled.

Currency
April 2026 accuracy

The book reflects Claude behavior as of April 2026, Claude 4.6. We are explicit about which details are likely to change and where to check for current information as the system evolves.