Agentic Engineering Maturity

Working well with AI is a craft with a learnable progression — not a question of whether you use AI (almost everyone does) but how. The Agentic Engineering Maturity (AEM) framework makes that progression explicit, so individuals, codebases, and teams can see where they are and what to practice next.

It was distilled from how the best practicing agentic engineers actually work — Boris Cherny (creator of Claude Code), Andrej Karpathy, Thariq Shihipar and the Anthropic engineering team, Simon Willison, Armin Ronacher, Geoffrey Huntley, Thorsten Ball, and others — cross-checked against established maturity-model literature (the Capability Maturity Model, DeepMind’s Levels of AGI, SAE autonomy levels, Factory.ai’s Agent Readiness, and DORA 2025).

---

The framework in one breath

• A 5-level Spine — your default mode of working with AI: Prompting → Prompt Engineering → Context Engineering → Agentic Engineering → Agentic Orchestration. Maturity is picking the right level for the task, not always the highest.
• 5 cross-cutting Axes — capabilities that deepen at every level: Verification, Context hygiene, Autonomy/Leash, Learning/Compounding, and Cost & Governance.
• 3 Scopes — Individual (workstation), Codebase (repo), Team (org).

The one rule that ties it together

Verification is the differentiator — the line between “vibe coding” and engineering. You can’t exceed Level 3 on the Spine with bottom-tier verification. You don’t climb to autonomy; you earn it through verification.

---

The five levels

L1 · Prompting

Ask and accept. You type a request and take what comes back. Great for throwaways; no repeatability or review.

L2 · Prompt Engineering

Reusable, structured asks. Roles, examples, format constraints. You review diffs. Still stateless — each prompt is an island.

L3 · Context Engineering

Build the system that feeds the model. A maintained CLAUDE.md, a curated tool set, deliberate context management. The hinge from prompt to system.

L4 · Agentic Engineering

Delegate, verify, stay in the loop. Agents act in a loop against a spec and a check; you review every diff. Quality is never optional.

L5 · Agentic Orchestration

Coordinate fleets, close the loops. Multiple agents, evals, and learning loops that improve without you re-engineering each cycle.

---

Explore the framework

The Maturity Model The Spine, the Axes, the Scopes, and the design lineage.

Pattern Library 24 reusable patterns from the best engineers, tagged by level and axis.

Self-Assessment Place yourself, your repo, and your team — with concrete signals.

Who to Learn From The practitioners worth following and how to track them.

Reading List The canonical posts, talks, and papers behind the framework.

---

Why this exists

AIOS is an open toolkit for working with AI in a structured, agent-native way. AEM is the part that answers “am I getting better at this, and how do I improve?” — a shared vocabulary and a set of practices anyone can adopt, whether you’re a solo developer, a founder, or a team lead.