AI Expert · 6th cohortin Spanish · starts November 3 · 40 seats · −33%

AI already writes codefaster than you.Learn to direct it

Build an agent-based workflow to research, develop and verify software with engineering judgment.

agent — your-projectwith harness

add Google sign-in

→ reading AGENTS.md · specs/ · DESIGN.md

→ spec: specs/add-google-sign-in.md · 3 acceptance criteria

→ worktree ../add-google-sign-in · implementer + validator

→ implementing only the files in scope

✓ validator: tests 11/11 · lint ok

✓ PR #145 ready for your review

Ideas:

Enter to run · try help or rm -rf /

developers trained
10,000+
years building software
15+
Google Developer Expert
GDE
of AI Expert, live
6th cohort
90 student testimonials
★★★★★

Where you may have seen us

GoogleBBVATelefónicaDeloitteNTT DataWallapopIdealistaGlobantSngularEventbriteCodemotionKairos
The problem

From loose prompts to a working system

The same request, two very different outcomes. The difference isn't the model or better prompting: it's the environment around the agent.

agent — with your harness

make me Google login

→ reading AGENTS.md and context

→ spec: scope + acceptance criteria

→ implementing (only files in scope)

✓ tests 12/12 · lint ok

✓ PR ready for your review

tests
✓ 12/12
scope
only what the spec says
your role
reviewing the PR

That environment has a name: harness engineering. It is what we teach, what we implement with teams and how we build software.

Harness engineering

Six layers that turn an agent into a working system

Models change every few weeks. What makes agent work reliable is the environment around them: the harness. Keep scrolling and watch it come together, layer by layer.

  1. layer 01 · research

    Discovery

    Before writing code, the agent researches with you. A discovery skill turns a fuzzy idea into an MVP with a clear scope.

    discovery — idea.md

    $ skill discovery "clinic bookings"

    → questions on users, limits and risks

    → research.md: competitors and constraints

    ✓ mvp.md: 5 features and what is out

  2. layer 02 · context

    Context engineering

    Rules, architecture and decisions live in the repo. Every session starts knowing where it is, without you repeating yourself.

    AGENTS.md

    # AGENTS.md

    - Domain free of framework dependencies

    - Specs in /specs before implementing

    - Nothing merges without green tests

    ✓ same context for people and agents

  3. layer 03 · execution

    Execution environments

    The tool does not matter. Hooks and subagents with separate roles: one implements, another validates, each with clear limits.

    agents — roles

    $ agent run implementer specs/export-csv.md

    → implementer: only files in scope

    → validator: read-only, against the spec

    ✓ pre-commit hook: lint + tests

  4. layer 04 · loop and verification

    Loop engineering

    Agents in parallel, each in its own worktree with its own PR. Work closes with evidence (tests, evals, Definition of Done), not promises.

    loop — feature_list

    $ git worktree add ../export-csv

    → 3 agents in parallel, no collisions

    ✓ tests 24/24 · evals ok

    ✓ PR #143 ready for your review

  5. layer 05 · product

    AI Product

    AI stops living only around your code and moves into your app: APIs, function calling, RAG and local models where they make sense.

    src/ai/rag.ts

    await rag.search("return policy")

    → 4 relevant chunks · embeddings

    POST /v1/chat/completions

    ✓ 200 OK · answer citing sources

  6. layer 06 · orchestration

    Your own agent

    All together: an orchestrated agent that moves features from spec to merge. You decide where to review, and nothing merges without you.

    orchestrator.ts

    ✏️ dark-mode → spec

    ⚙️ search-filter → implementing

    👀 auth-login → waiting for your review

    ✅ 12 merged · human supervision

✓ harness complete

The agent is whatever comes next. The harness is yours.

It lives in your repo, works with Codex, Claude Code, Cursor or OpenCode and does not expire when the model changes. It is exactly what you build in AI Expert.

Build your harness in AI Expert (in Spanish)
Testimonials

What changes when you stop asking and start directing

Their own words, as a diff: in red how they worked before, in green what they took away. Nothing rewritten, only trimmed.

review — ai-expert✓ 6 · approved

diff --git a/alumnos/santiago-perez-barber.md b/alumnos/santiago-perez-barber.md

@@ AI Expert · +19 −8 @@

1You start out just "tinkering" with AI tools…

2+…end up mastering a set of resources that are useful for your daily work…

3+…and exponentially increase your productivity.

Santiago Pérez BarberBytacora Soluciones Informáticas SL · ★★★★★

And the rest of the history

90 reviews · ★★★★★ · every program

✓ approved · ★★★★★

Lab

The latest things we have tried with agents

We test tools, models and workflows on real projects and share what works and what does not. No hype.

Go to the blog
zsh — devexpert

$git log --oneline blog/

$ git show 87e8012

commit 87e8012

Author: Antonio Leiva · 2026-07-08 · 11 min

People are starting to talk quite a lot about Loop Engineering. As always happens when a new label appears, part of it is hype, but another part is genuinely useful: once something has a name, we can discuss it more…

$

Instructors

Instructors who apply AI in real projects

Antonio and Nino combine software development, agent systems, automation, data and teaching experience. They work directly on both training and company projects.

tmux · devexpert
Antonio Leiva

antonio@devexpert

Antonio LeivaGDE

DevExpert founder · Google Developer Expert

Nino Ruano

nino@devexpert

Nino RuanoAI · Big Data

AI and Big Data specialist · Instructor

It is a real shell (well, almost). Type help to see the commands.

Next step

Stop fighting the agent. Start directing it.

Learn with us in AI Expert or tell us what your team needs. You build the harness once; what changes is everything you can do with it.

$ hook pre-tool --interactive

blocked 0lives ●●●time 00:00best 0

main · production

$ hook pre-tool --interactive

Hook Defender

You are the hook watching what the agent runs. Block destructive commands before they reach production and let the good ones through.

  • → Click or tap a command to block it.
  • → If a destructive command reaches main, you lose a life.
  • → If you block a safe one, the agent gets stuck: you lose one too.