The 10 Levels of Agentic Engineering — Where Are You?

Where Are You on the Agentic Engineering Spectrum?

AI-assisted coding is not new. But the gap between engineers who use AI as a search engine and those who orchestrate autonomous agent swarms is enormous and growing every week. Most engineers sit somewhere in the middle, unsure how to level up or even how to describe what "good" looks like.

This framework gives you a unified scale. Ten levels, from zero to nine, each with clear markers so you can honestly assess where you are and see exactly what the next step looks like.

The Framework

Each level builds on the one before it. You cannot meaningfully operate at Level 5 without having the foundations of Levels 0 through 4 in place. The levels are cumulative, not alternative.

L0 — Terminal Tourist

You type prompts into an AI tool. You have no configuration, no context files, no custom instructions. Every conversation starts from scratch. You copy-paste code out of chat windows and hope it works.

• No rules file or context configuration
• No memory between sessions
• Manual copy-paste workflow
• AI is a search engine replacement

Most engineers start here. There is no shame in it, but there is no reason to stay.

L1 — Grounded

A rules file exists. You have told the AI about your stack, your standards, and your project structure. The AI now has baseline context before you type a single prompt.

• CLAUDE.md, .cursorrules, or copilot-instructions.md is configured
• Stack, language, framework, and conventions are documented
• Project structure is described
• The AI produces relevant code on the first try more often

L2 — Connected

Your AI agent is connected to real data sources. MCP servers, database connections, API documentation, and live systems feed context into the agent automatically.

• MCP servers configured and pulling real data
• Agent can query databases, read documentation, access GitHub
• Context is dynamic, not just static files
• The agent understands your actual system state, not just your description of it

L3 — Skilled

You have three or more custom skills. These are repeatable, documented workflows that the agent can execute without re-explanation every time.

• 3+ custom skills (SKILL.md files or equivalent)
• Skills cover repeatable workflows: PR review, deployment, testing, documentation
• Agent can execute multi-step tasks from a single command
• Workflows are consistent and auditable

L4 — Compounding Architect

You operate in a continuous loop: Plan, Delegate, Assess, Codify. Every task the agent completes feeds learnings back into the system, making the next task faster and better.

• Plan: Break down work before delegating
• Delegate: Assign to the right agent or skill
• Assess: Review output, identify gaps
• Codify: Update context, skills, and knowledge with what you learned

This is where the compounding effect begins. Your system gets smarter with every interaction.

L5 — Harness Builder

You have built automated feedback loops that create backpressure on the agent. Tests, linters, type checkers, and pre-commit hooks act as guardrails that force the agent to self-correct.

• Tests run automatically after agent generates code
• Linters and formatters enforce standards without human review
• Pre-commit hooks catch issues before they reach the repository
• The agent iterates until the harness passes, not until it "looks right"

L6 — Pipeline Engineer

The AI runs in headless mode. Scripts call the agent, not humans. CI pipelines trigger agent workflows. The agent is part of your infrastructure, not just your IDE.

• Headless / non-interactive mode configured
• Scripts and CI jobs invoke the agent programmatically
• Agent runs as part of deployment, testing, or review pipelines
• Human involvement is supervisory, not operational

L7 — Multi-Agent Operator

Multiple agents run in parallel, coordinated by an orchestrator. Different models handle different jobs based on their strengths and cost profiles.

• Parallel agents working on independent tasks simultaneously
• Orchestrator coordinates task assignment and result collection
• Different models for different jobs (reasoning vs coding vs reviewing)
• Cost-aware routing: expensive models for planning, cheap models for execution

L8 — Always On

Agents run without you. Cron jobs, background processes, and cloud VMs keep agents working around the clock. You wake up to completed tasks, PRs ready for review, and reports waiting in your inbox.

• Cron-triggered agent workflows
• Background agents monitoring and responding to events
• Cloud VMs running persistent agent processes
• Asynchronous notification when work completes

L9 — Swarm Architect

Agents manage other agents. The system creates, assigns, monitors, and retires agents dynamically. Direct coordination between agents replaces human orchestration.

• Agents spawn and manage sub-agents
• Dynamic task decomposition and assignment
• Agent-to-agent communication and coordination
• Self-healing: failed agents are replaced automatically
• Human role shifts from operator to architect

How to Use This Framework

Be honest about your current level. Most engineers who think they are at L4 are actually at L1 or L2. That is fine. The framework is not a competition; it is a map.

Practical steps:

• Assess yourself now. Read each level and identify where you genuinely operate today.
• Focus on the next level only. Do not try to jump from L1 to L6. Each level builds foundations the next level requires.
• Measure by output, not tooling. Having MCP configured does not make you L2 if you never use it. Levels are about practice, not installation.
• Reassess monthly. The field moves fast. What was L7 six months ago may be standard practice soon.

Conclusion

Agentic engineering is not a buzzword. It is a measurable skill set that determines how much leverage you get from AI tools. The difference between L0 and L9 is not intelligence or experience; it is deliberate investment in context, skills, automation, and architecture.

Start where you are. Move one level at a time. The compounding effects are real.

This framework draws on ideas from Bassim Eledath and the broader agentic engineering community.