Learning roadmap
AI Engineering Architect
From AI-assisted development to AI-native engineering leadership.
A practical, structured learning journey to master AI agents, context engineering, harness engineering, evaluation frameworks, production deployment, and AI-driven software development. For experienced software engineers, technical founders, engineering managers and technology leaders moving into AI-native engineering.
- 10 weeks
- About 102 hours
- Intermediate to advanced
- Hands-on projects
- 83 official OpenAI and Anthropic resources
What you will learn
- 01AI architecture & solution design
- 02Context engineering
- 03Agent skills & tool development
- 04MCP servers & integrations
- 05Agent harness engineering
- 06Evaluations & reliability
- 07AI-native SDLC
- 08Production deployment & governance
- 09Forward deployed engineering
- 10Business value & ROI
The roadmap
Ten weeks, one deliverable each.
- 010/8
Claude Code, Codex, project configuration, CLAUDE.md, AGENTS.md, permissions.
Beginner · 8 h · 7 links
Deliverable: A repository configured for two coding agents, with instructions and a permission policy.
7 links
Beginner8 h
A repository configured for two coding agents, with instructions and a permission policy.
Courses and reading
- ~2 h
Anthropic Academy · Course
Claude Code 101 (opens in a new tab) - ~15 min
Claude Code Docs · Documentation
How Claude remembers your project (opens in a new tab) - ~20 min
Claude Code Docs · Documentation
Configure permissions (opens in a new tab) - ~10 min
AGENTS.md · Documentation
AGENTS.md (opens in a new tab) - ~10 min
OpenAI Codex Docs · Documentation
Custom instructions with AGENTS.md (opens in a new tab) - ~15 min
OpenAI Codex Docs · Documentation
Agent approvals & security (opens in a new tab) - ~1 h
OpenAI Academy · Course
Codex Fundamentals - Video (opens in a new tab)
Assignment
Pick an existing service. Give it a CLAUDE.md and an AGENTS.md that describe how to build, test and follow its conventions, and a permission policy that allows the test suite but blocks secrets and production credentials. Complete the same three tasks with each agent (a bug fix, a small feature and a refactor) and log where each one needed correcting.
Deliverable: A repository configured for two coding agents, with instructions and a permission policy.
Done when
- Both instruction files are committed and kept short enough to read in two minutes
- Agents run the tests without asking, and are refused on deny-listed files
- A one-page comparison of the six runs: time taken, corrections needed, final quality
You will learn to
- Run two terminal coding agents productively on a real codebase.
- Write project instructions that agents actually follow.
- Choose permission settings that match the risk of each action.
Before you start
- Professional experience with git and at least one backend language
- A repository you are free to experiment on
- Access to Claude Code and Codex through a paid plan or an API key
- 020/9
Context management, custom skills, reusable instructions, tool selection.
Intermediate · 8 h · 8 links
Deliverable: Two reusable skills built from real team workflows, with before-and-after results.
8 links
Intermediate8 h
Two reusable skills built from real team workflows, with before-and-after results.
- 030/9
Model Context Protocol, APIs, tool development, authentication, hooks.
Intermediate · 10 h · 8 links
Deliverable: A working MCP server for a real API, with authentication and a guarding hook.
8 links
Intermediate10 h
A working MCP server for a real API, with authentication and a guarding hook.
- 040/11
Agent design patterns, orchestration, tool calling, workflows vs agents.
Intermediate · 10 h · 10 links
Deliverable: An architecture decision record comparing a workflow and an agent on the same task.
10 links
Intermediate10 h
An architecture decision record comparing a workflow and an agent on the same task.
- 050/12
Agent execution loops, checkpoints, error recovery, sandboxing, observability.
Advanced · 12 h · 11 links
Deliverable: A long-running agent harness with checkpoints, recovery, a sandbox and traces.
11 links
Advanced12 h
A long-running agent harness with checkpoints, recovery, a sandbox and traces.
- 060/8
Evaluation datasets, automated graders, reliability testing, regression detection.
Advanced · 10 h · 7 links
Deliverable: An evaluation suite that runs in CI and catches a planted regression.
7 links
Advanced10 h
An evaluation suite that runs in CI and catches a planted regression.
- 070/10
Prompt optimization, model selection, cost, latency, reliability.
Advanced · 8 h · 9 links
Deliverable: A before-and-after report on quality, cost and latency, guarded by your evals.
9 links
Advanced8 h
A before-and-after report on quality, cost and latency, guarded by your evals.
- 080/9
AI-assisted development workflows, automated code review, testing, GitHub Actions, CI/CD.
Intermediate · 10 h · 8 links
Deliverable: A CI pipeline where agents review pull requests and draft fixes, with humans approving.
8 links
Intermediate10 h
A CI pipeline where agents review pull requests and draft fixes, with humans approving.
- 090/12
Security, deployment, monitoring, permissions, governance, rollback.
Advanced · 10 h · 11 links
Deliverable: A production readiness review, runbook and tested kill switch for your agent.
11 links
Advanced10 h
A production readiness review, runbook and tested kill switch for your agent.
- 100/5
Design and implement an end-to-end AI engineering workflow.
Advanced · 16 h · 4 links
Deliverable: A working end-to-end workflow, a recorded demo and a decision-ready write-up.
4 links
Advanced16 h
A working end-to-end workflow, a recorded demo and a decision-ready write-up.
Hours are our estimates for study plus the hands-on work, and vary with experience.
Week 10 · Capstone
Build the whole workflow.
01Business requirement
A plain-language request enters the system as an issue or ticket.
02AI planning
An agent proposes a plan, the files it will touch and the tests it will add.
03Implementation
A coding agent implements the plan in a sandbox, inside its harness.
04Automated testing
Unit and integration tests run, and failures go back to the agent.
05Evaluation
Your evaluation suite scores the change against the quality floor.
06Code review
An automated reviewer comments, and the agent addresses what is valid.
07Human approval
A named person approves the plan, the diff and the evidence.
08Deployment
The change ships behind a flag or canary, with monitoring and rollback ready.
Acceptance criteria
- Runs end to end on three real requirements with no manual steps other than approval
- Every stage leaves a trace: who or what acted, with inputs, outputs and cost
- A failing test or evaluation stops the pipeline before review
- Nothing reaches production without a recorded human approval
- A bad deploy can be rolled back in one command, and this has been tested
- Secrets never enter the model's context, and permissions are least-privilege
- Another engineer can set the workflow up on a new repository from its README
Success metrics
| Metric | Suggested target |
|---|---|
| Lead time, requirement to approved change | At least 50% shorter than your measured baseline |
| Evaluation pass rate | At or above the floor set in week 7 |
| Changes approved without rework | At least 70% |
| Cost per completed change | Measured and reported, with a stated budget |
| Unapproved production changes | Zero |
| Time to roll back | Under 10 minutes, rehearsed |
Targets are suggestions. Set your own baseline in week 1 and agree the targets before you start the capstone.
Skills assessment
Rate yourself honestly.
Readiness 0%
Context engineering
Skills development
MCP
Harness engineering
Evals
AI architecture
AI-native SDLC
Production engineering
Technical leadership
An independent curriculum published by ZL Tech. It is not affiliated with, sponsored by or endorsed by Anthropic or OpenAI, and completing it is not a certification. Course names and links belong to their providers and may change. Last reviewed October 2026.