Agentic Levels

Original article

Bassim Eledath

Original "8 Levels of Agentic Engineering", the framework this entire site is built on

What shaped the framework

Thinkers, tools, and projects whose ideas are woven into the level model.

Claude Code, Claude models, and the MCP protocol that make Level 4–8 possible

Shipped Claude Code and the Model Context Protocol, creating the infrastructure that makes Levels 4 through 8 practically achievable.

Geoffrey Huntley

The Ralph loop: a practical blueprint for autonomous background agent execution

The Ralph loop gave Level 6+ practitioners a repeatable pattern for autonomous background execution.

HumanLayer / 12-Factor Agents

12 design principles for reliable agent software, referenced at every level

Provided the clearest architectural vocabulary for building reliable agent software, referenced across every level of the framework.

Andrej Karpathy

Autonomous research loops, synthesis patterns, and literature-driven agent work.

Demonstrated that autonomous research loops could produce synthesis-quality output, validating the harness engineering approach at Level 6.

Microsoft AutoGen

One of the clearest reference points for production-style multi-agent orchestration and coordination patterns.

Set a concrete benchmark for multi-agent orchestration that practitioners at Levels 7 and 8 measure against.

People we learned from

Writers and practitioners who published the patterns that became this framework's vocabulary.

Kevin Gargate Osorio + Frank Andrade

Their thinking on separating role, rules, boundaries, context, and skills instead of overloading one prompt shaped how this site explains CLAUDE.md design.

Introduced the structural separation of role, rules, and context that became foundational to CLAUDE.md design.

WAVIBOY

Showed that creative agents can work in parallel lanes, each with defined roles, handoffs, and explicit outputs, without one model trying to do everything.

Demonstrated that creative teams could structure agent workflows with explicit handoffs and parallel execution lanes.

Vikas Sah

Made atomic skill design and versioned skill libraries practical for individual practitioners, not just engineering teams.

Made skill versioning and atomic skill design accessible to people at Level 5, accelerating custom capability development.

Gabor Meszaros

CLAUDE.md formatting patterns that improve agent compliance through rationale, hierarchy, and naming.

Established concrete formatting rules for CLAUDE.md that measurably improve how faithfully agents follow project conventions.

@keshavsuki

The layered .claude folder blueprint and practical memory-system patterns.

The layered .claude folder blueprint gave people a concrete structure for managing memory and context across long projects.

Ruben Hassid

Skills as reusable team workflows, plus evaluation-first adoption instead of one-off prompt improvisation.

Reframed skills as team-level reusable workflows, shifting the conversation from individual prompts to shared organizational assets.

Maverick

Sabrina Ramonov

Made CLAUDE.md, plan mode, and context hygiene approachable for people who had never used an AI coding tool before.

Made CLAUDE.md and plan mode accessible to people who had never used an AI coding agent before, lowering the entry barrier to Level 2.

Ewan Mak

A rigorous breakdown of the Claude Code harness: what it actually does, where it falls short, and what that means for long-term adoption.

Provided a rigorous analysis of the Claude Code harness that helped people understand what they were actually adopting and where it fell short.

Minervee

Role-reversal prompting and adversarial self-critique for better reasoning under stress.

Introduced adversarial self-critique as a practical technique that improved AI output accuracy without requiring more capable models.

Rick Hightower

Context hygiene patterns with /btw, /fork, and /rewind for cleaner long-running sessions.

Codified context hygiene commands into a repeatable toolkit that extended the useful life of long agent sessions.

Shan Rai Shah

Practical Claude Code best-practice patterns covering repeatable rules, memory habits, and cleaner session setup.

Published a battle-tested reference that people could adapt, cutting the time to a working CLAUDE.md from days to hours.

SuperClaude Org

An open-source Claude Code operating system that shows what a structured, scalable agent environment looks like in practice.

Built a full Claude Code operating system framework that showed what a structured agent environment looks like at scale.

Communities we hang out in

Places where people doing this work share what they're learning.

Agent Architects

Where people building agent systems share architecture patterns, workflow breakdowns, and real operating decisions.

Provided a community where real agent architecture decisions are discussed and refined outside of vendor documentation.

AI Staffroom

Where early practitioners share workflow experiments, operating tips, and honest takes on what actually works.

Surfaced real-world workflow experiments and operating tips that shaped the practical guidance throughout Levels 0–4.

Quantum Quill Lyceum

A learning-first space where experimentation is the norm and shared resources make it easier to move through the early levels.

Cultivated a learning-first environment where experimentation and shared resources helped members advance through the early levels faster.

Chase AI Community

Collects tactics, examples, and operator workflows across multiple AI tools from people actively using them.

Collected and shared operator workflows across multiple AI tools, helping people build tool-agnostic habits at the foundation levels.

Where this came from

Bassim Eledath

Where this came from

Bassim Eledath