# How To Become An AI Engineer in 2026 > Without a CS degree. Without a bootcamp. Without knowing what a transformer is today. [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE) [![Oriro-Labs](https://img.shields.io/badge/Oriro--Labs-Learning%20Series-blue)](#) [![2026](https://img.shields.io/badge/Updated-June%202026-orange)](#) [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](CONTRIBUTING.md) --- ## The One Number You Need to Know ``` Same model. Two harnesses. CORE benchmark: Claude Code harness: 78% Smolagents harness: 42% ─── 36-point gap ``` **The harness is the job.** This repository teaches you how to build it. --- ## What's In Here ``` become-an-ai-engineer-26/ │ ├── SKILL.md ← The complete engineering skill (start here) ├── BecomeAnAIEngineer26.pdf ← Illustrated PDF with all diagrams ├── README.md ← This file ├── CONTRIBUTING.md ← How to contribute ├── LICENSE ← MIT │ ├── phases/ │ ├── phase-0-mental-models/ │ │ ├── README.md ← Workflow vs Agent, 5 patterns, harness model │ │ └── project/ │ │ └── TEMPLATE.md ← 2-page concept doc template │ │ │ ├── phase-1-first-agent/ │ │ ├── README.md ← Raw loop + Agent SDK walkthrough │ │ └── code/ │ │ ├── raw_loop.py ← ~100-line raw agent loop │ │ └── agent_sdk_agent.py ← Same agent on Claude Agent SDK │ │ │ ├── phase-2-architecture/ │ │ ├── README.md ← LangGraph + Deep Agents │ │ └── code/ │ │ ├── research_agent/ ← Lead + 3 sub-agents + writer │ │ │ ├── graph.py │ │ │ ├── nodes.py │ │ │ └── tools.py │ │ └── requirements.txt │ │ │ ├── phase-3-harness/ │ │ ├── README.md ← 10 components walkthrough │ │ └── code/ │ │ └── mini_harness/ ← 1,500-line reference implementation │ │ ├── harness.py │ │ ├── tools.py │ │ ├── hooks.py │ │ ├── context.py │ │ └── persistence.py │ │ │ ├── phase-4-evals/ │ │ ├── README.md ← 4 eval types + CI setup │ │ └── code/ │ │ ├── evals/ │ │ │ ├── golden_dataset.py │ │ │ ├── trajectory_eval.py │ │ │ ├── llm_judge.py │ │ │ └── end_state_eval.py │ │ └── .github/ │ │ └── workflows/ │ │ └── eval.yml ← GitHub Actions CI gate │ │ │ └── phase-5-production/ │ ├── README.md ← Cost, latency, safety, monitoring, resilience │ └── code/ │ ├── cost_routing.py │ ├── sandboxing.py │ └── monitoring.py │ ├── projects/ │ ├── 1-mobile-app-slm/ ← Beginner: Edge AI + resource constraints │ ├── 2-self-improving-coder/ ← Intermediate: Agentic loops │ ├── 3-video-editor-agent/ ← Advanced: Multimodal AI │ ├── 4-personal-life-os/ ← Expert: Deep context + privacy │ └── 5-enterprise-workflow/ ← Master: Production orchestration │ ├── references/ │ ├── yarn-context-extension.md ← YaRN: 128K context in 10× fewer steps │ ├── mhc-stable-training.md ← mHC: Manifold-constrained residual streams │ ├── benchmark-numbers.md ← May 2026 benchmark scores │ └── stack-decisions.md ← Framework comparisons and decisions │ └── community/ ├── BUILDS.md ← Community project showcase └── DISCUSSIONS.md ← Open questions and ongoing research ``` --- ## The 17-Week Roadmap | Phase | Weeks | Build | Milestone | | --------------------- | ------- | ------------------------ | ------------------------------------- | | 0 — Mental Models | 1-2 | 2-page concept doc | Explain a harness in plain English | | 1 — First Agent | 3-5 | Raw loop + SDK agent | Agent shipped with Skill + hook | | 2 — Architecture | 6-9 | LangGraph research agent | Architecture solid + LangSmith traces | | 3 — Build the Harness | 10-13 | 1,500-line mini-harness | Harness complete + post-mortem | | 4 — Evals & CI | 14-17 | Regression harness | Eval suite in GitHub Actions | | 5 — Production | Forever | Harden everything | Never ends | **Moonlighting?** Multiply every duration by 2.5×. --- ## Start Here (Technical Quick-Start) ### Phase 0 (this weekend, no code) Read `SKILL.md` sections: Core Insight → What an AI Engineer Does → Phase 0. Write the 2-page doc. Do not skip this. ### Phase 1 (Weeks 3-5) ```bash cd phases/phase-1-first-agent/code pip install anthropic python raw_loop.py # understand the loop python agent_sdk_agent.py # feel what the harness gives you for free ``` ### Phase 2 (Weeks 6-9) ```bash cd phases/phase-2-architecture/code/research_agent pip install -r requirements.txt # Requires: ANTHROPIC_API_KEY, POSTGRES_URL, LANGSMITH_API_KEY python -m research_agent.main "What are the current AI engineering best practices?" ``` ### Phase 3 (Weeks 10-13) ```bash cd phases/phase-3-harness/code/mini_harness pip install -r requirements.txt python harness.py --task "Research and summarise the top 5 AI frameworks in 2026" ``` ### Phase 4 (Weeks 14-17) ```bash cd phases/phase-4-evals/code python evals/run_golden.py --dataset evals/golden_dataset.jsonl python evals/trajectory_eval.py --traces traces/ # Copy .github/workflows/eval.yml into your repo to gate CI ``` --- ## The 5 Portfolio Projects Pick one. Build it this weekend. Ship something real. | # | Project | Level | What It Proves | | --- | --------------------------------------------------------------- | ------------ | ------------------------------- | | 1 | [Mobile App + SLM](projects/1-mobile-app-slm/) | Beginner | Edge AI, resource constraints | | 2 | [Self-Improving Coding Agent](projects/2-self-improving-coder/) | Intermediate | Agentic loops, debugging | | 3 | [Video Editor Agent](projects/3-video-editor-agent/) | Advanced | Multimodal AI, tool integration | | 4 | [Personal Life OS Agent](projects/4-personal-life-os/) | Expert | Deep context, privacy-first | | 5 | [Enterprise Workflow Agent](projects/5-enterprise-workflow/) | Master | Production orchestration | --- ## The Stack Decision ``` ✅ Use: LangGraph 1.0 + Deep Agents (state machine + durability + observability) ✅ Use: Claude Agent SDK (same harness as Claude Code) ✅ Use: LangSmith / Braintrust / Phoenix (pick one for observability) ❌ Avoid: CrewAI in production (fast demo, fragile at scale) ❌ Avoid: OpenAI Swarm (their README says "not production-ready") ❌ Avoid: OpenAI Assistants API (sunsetting mid-2026) ⚠️ Demo: CrewAI (fine for hackathons, not production) ``` --- ## Benchmark Context (May 2026) The leaderboard changes weekly. What doesn't change: ``` Same model, different harness = 10-36 point swing on any benchmark. SWE-bench Verified: Claude Opus 4.7 ~87.6% | GPT-5.5 ~88.7% GAIA (agents): Claude Sonnet 4.5 74.6% ← leads τ-bench (service): Claude Mythos Preview 89.2% ← leads ``` **The engineer who builds the harness writes the benchmark score.** --- ## Advanced References For engineers who want to go deeper: - **[YaRN: Context Window Extension](references/yarn-context-extension.md)** — How to give a model 128K token context using 10× fewer training steps than prior methods. ICLR 2024. - **[mHC: Stable Training at Scale](references/mhc-stable-training.md)** — How DeepSeek solved training instability at 27B+ parameters using Sinkhorn manifold constraints. arxiv:2512.24880. - **[Benchmark Numbers](references/benchmark-numbers.md)** — Current leaderboard with context on what each benchmark actually measures. - **[Stack Decisions](references/stack-decisions.md)** — Detailed framework comparison with reasoning for each recommendation. --- ## Contributing This is a living document. The field moves weekly. ``` What we want: ✓ Phase project implementations (open PRs) ✓ Bug fixes in example code ✓ Benchmark updates as new scores publish ✓ New reference material on emerging techniques ✓ Translations ✓ Your build in BUILDS.md What we don't want: ✗ Promotional content ✗ Framework comparisons without benchmark evidence ✗ "AI wrapper" projects presented as agent systems ``` See [CONTRIBUTING.md](CONTRIBUTING.md) for the full guide. --- ## The Uncomfortable Truth Most people who read this will do nothing. They will bookmark it, say "great resource," and go back to building wrappers. 57% of teams now have agents in production. 89% of those have observability wired. Quality is the #1 barrier — 32% of teams cite it as the blocker. **The entire field is bottlenecked on engineers who can build evals and harnesses. Not on engineers who can call an LLM API. That is the job opening.** --- ## License MIT — use freely, build openly, teach everyone. --- _AI Engineering by · Oriro-Labs Learning Series · 2026_