Agentic engineering
A practical curriculum for developers and technical builders learning to design, evaluate, and operate reliable AI agents.
Developers and technical builders who need to understand the whole agent system—not just prompting or one tool.
From choosing a model to operating a multi-agent system
Each level adds a responsibility: understand the model, direct the work, supply context, enable action, verify behavior, operate reliably, then coordinate at scale.
Learn the software beneath the agent
Learn enough shell, Git, APIs, databases, testing, builds, and deployment to direct and verify an agent's work.
Open on-ramp →- 01Model literacy
Understand the model
Choose models for real workloads and understand their limits, costs, and serving options.
Open level → - 02Intent and control
Direct the work
Turn goals into instructions, constraints, outputs, and approval points an agent can follow.
Open level → - 03Context and retrieval
Engineer context and knowledge
Give agents the right instructions and knowledge without overflowing the context window.
Open level → - 04Tools and execution
Give the agent capabilities
Connect agents to tools, protocols, browsers, sandboxes, and typed outputs so they can take bounded action.
Open level → - 05Evals and verification
Close the feedback loop
Verify behavior with specifications, tests, evals, traces, threat models, and feedback loops.
Open level → - 06Production systems
Operate reliable agents
Add identity, data, isolation, observability, budgets, and recovery for real users and long-running work.
Open level → - 07Orchestration
Compound and scale
Coordinate multiple agents only after decomposition, isolation, verification, and review are reliable.
Open level →
Choose a level
Begin at Level 1, take the optional foundations, or jump to a responsibility you already own.
Learn a concept
Read the explanation with current evidence, relevant tools, and the surrounding concepts in one place.
Check understanding
Mark what you understand and use each level assessment to find the next gap.
Learn the system, not a stack of products
Lessons stay open without an account. Sign in only to save progress and assessment results.
Editorial notes and methodology
The maintenance agent audits definitions, prerequisite relationships, attached Intel, and source freshness. Levels remain an editorial framework and change only when evidence shows the sequence no longer reflects how reliable agent systems are built.
Reviewed every 14 days
18 evidence gaps remain visible to maintainers. The current snapshot covers 278 Intel items.
Roadmap changelog
- Edition 2026.8
Added the missing software prerequisites and decision lessons, then completed the multimodal, agent-experience, security, and data-governance production curriculum.
- Edition 2026.7
Replaced every remaining definition-and-link fallback with a complete guided lesson and added inline media where watching teaches the concept better.
- Edition 2026.6
Extended the guided lesson reader, evidence composition, tool paths, and sequential navigation across all seven roadmap levels.
- Edition 2026.5
Turned Model Literacy into an addressable guided reading sequence and hardened every roadmap citation against local Intel drift.
- Edition 2026.4
Expanded the roadmap into a 70-concept agentic-engineering progression with live evidence, tool paths, and server-scored level assessments.
Bassim Eledath — The 8 Levels of Agentic Engineering
Context, compounding, capabilities, harnesses, background execution, and the limits of autonomous teams.
Steve Yegge's 8 Levels of AI Development
The shift from synchronous assistance to CLI work, parallel agents, and orchestration—and the verification burden that follows.
roadmap.sh — AI Engineer
Dense concept navigation and broad coverage of model fundamentals, retrieval, agents, safety, and production tooling.