Ten concrete proposals for systems engineering education in Mexico now that AI writes code

By Alfredo Castillejos · Dixi-Project · November 11, 2026 · 9 min read
Illustration: a list of ten checkboxes, the first three checked in green, next to a shield with a check mark.

In What to study (and what not to) in a systems engineering degree for 2032 we proposed a complete 9-semester curriculum. A full curriculum is useful for thinking, but it takes years to approve. This article answers a shorter question: if we had to propose today, point by point, what should change in systems engineering education in Mexico, what would we propose?

The answer starts from one fact and one risk.

  • The fact: writing an implementation from a description no longer sets an engineer apart. Current models do it in minutes, and 18.7% of the PRs merged on GitHub in September 2026 were signed by AI agents (verification debt).
  • The risk: employment of 22- to 25-year-olds in the occupations most exposed to AI fell 19% below that of their less-exposed peers (Brynjolfsson, Chandar and Chen, 2026). This is the broken pipeline: if nobody gets in, in ten years there will be nobody to supervise the AI.

As AI produces more and more code, what will be scarce in Mexico is people who define the problem, verify what the machine delivers, run the system in production and answer for the result. The ten proposals below aim to train those people, ordered by impact and by how quickly they can be done.

The ten proposals at a glance

# Proposal Who does it Requires a curriculum reform?
1 Require evidence of verification for accreditation CACEI and COPAES No
2 Internships and social service as audits of real systems Each institution No
3 Train faculty before changing the paperwork TecNM, universities, ANUIES No
4 Redesign the specialization now Each institution, with its approval Only the specialization
5 Do not cut mathematics or shorten the degree Curriculum designers Yes
6 A mandatory course on Mexico's regulatory context Curriculum designers Yes
7 Tools in quarterly labs, not in courses Each institution No
8 Shared compute and open models for public universities ANUIES and SECIHTI No
9 Rebuild the entry door for people starting out Industry and federal government No
10 An accountable human for critical public systems Government and professional associations No

Seven of the ten do not require touching a curriculum. That is deliberate: the TecNM Computer Systems Engineering curriculum has been in force since 2010, and waiting for a full reform means letting another generation go by.

First, what does not require a curriculum reform

1. Require evidence of verification to accredit a program

The fastest lever is not creating courses but changing what is required to accredit an engineering program. If CACEI, the COPAES-recognized body that accredits engineering programs in Mexico, asked for these three pieces of evidence, every school seeking accreditation would adopt them at once:

  • Dual assessment: what students do without AI (individual fundamentals) is graded separately from what they do with AI (project).
  • Seeded-defect audit: students receive AI-generated code with known bugs, and we measure how many they find. It is graded automatically, so it scales to groups of 120.
  • Oral defense: every student explains any part of what they submit.

None of the three changes a course on paper. They change what counts as evidence that someone learned.

2. Turn internships and social service into audits

Today the professional internship (residencia profesional) often ends with "I built a small system for a company". We propose a different main deliverable: specifying, testing and auditing real systems for municipalities, public hospitals, schools and small businesses.

The benefit is twofold. More than 99% of Mexico's business establishments are micro, small or medium-sized (INEGI, Economic Censuses), and almost none can afford a security review or a test suite. Students learn what the market will ask for, and the organization gets a review it would not otherwise have.

3. Train faculty before changing the paperwork

Any reform that asks to teach verification, security and development with AI agents fails if the faculty have not practiced them. We propose one year of faculty training before any formal change:

  • Testing and verification, including mutation testing.
  • Security and threat modeling.
  • Development with AI agents under controls: tests first, risk-based review, an AI usage log.
  • Designing assessments with seeded defects.

Without this, the change stays in the course names.

In the curriculum

4. Redesign the specialization now

At TecNM, specialization courses are approved much sooner than the common core. That is the place to start. An immediate version, with five courses:

  • Machine learning and evaluation
  • Foundation models
  • Data engineering
  • AI systems engineering
  • Evaluation and security of AI systems

5. Do not cut mathematics or shorten the degree

If AI writes the code, the temptation is to drop mathematics and graduate in fewer semesters. Both are bad ideas:

  • Probability and inference are what let you tell whether a model works or only seems to. Without them, AI becomes a black box nobody can audit.
  • Shortening the degree sends less-trained people into a market that hires fewer juniors and demands more judgment from them. Besides, AI without guardrails can harm learning: students who practiced with unrestricted AI scored 17% lower on the exam than those who never used it (Bastani et al., PNAS, 2025).

6. A mandatory course on Mexico's regulatory context

A general-purpose model does not know Mexico's legal and tax context in depth. An engineer who masters it is hard to replace. The course would cover, at a minimum:

  • The Federal Law on the Protection of Personal Data Held by Private Parties published in 2025.
  • Electronic invoicing (CFDI) and obligations to the tax authority (SAT).
  • Health regulations and official standards (NOMs) that affect information systems.
  • Government procurement and public systems.

This course also builds something AI cannot do for a person: understanding the client's domain.

7. Tools in quarterly labs, not in courses

No course called "Prompt engineering" and none named after a framework. Tools change every quarter and a formal curriculum takes years to change. We propose three layers:

Layer What it contains Reviewed
Core Fundamentals, verification, security, domain and accountability Every 4 or 5 years
Competencies The learning outcomes of each course Every year
Labs Specific models, agents, frameworks and cloud services Every quarter

When a new generation of models appears, the lab is updated, not the curriculum.

Public policy and industry

The last three proposals are public policy. They are our opinion, not something measured like the data above.

8. Shared compute and open models for public universities

If labs depend on paid licenses, the institutes with fewer resources are left out. It is the biggest equity risk of any reform. We propose a national consortium (for example, ANUIES with SECIHTI, Mexico's science ministry) offering:

  • A compute budget per student.
  • Open models hosted in Mexico, to practice without sending data abroad.
  • Labs shared across campuses.

9. Rebuild the entry door for people starting out

Companies hire fewer juniors to program, because AI already does that. The answer is not to lament it but to create a different entry door: apprentice roles in verification, testing, operations and security, where the person reviews what AI produces under the supervision of someone experienced.

Some existing instruments could be pointed at this, such as a tax incentive per apprentice or a track within Jóvenes Construyendo el Futuro, the federal on-the-job training program for people aged 18 to 29. If nobody gets in today, in ten years there will be no seniors to supervise the AI.

10. An accountable human for critical public systems

For health, finance, registry and security systems, we propose that a licensed engineer (cédula profesional) sign off on the system, with a record of what the AI did and how it was verified. It answers the most worrying question about AI-written software: who answers for it?. It also gives the profession weight.

What we do not propose

  • Short "applied AI" programs that only teach how to use tools. They expire before the first cohort graduates.
  • Banning AI. Separate where it is used and where it is not; do not ban it.
  • Measuring graduates by how many CRUD apps they deliver. A working CRUD app proves nothing anymore. Better to publish how many graduates are employed after 12 months, and in what role.

If only one thing could be done

Grade separately what students do without AI and with AI, with an oral defense and a seeded-defect audit. It is cheap, needs no curriculum reform, scales to large groups because the audit grades itself, and forces everything else: to pass, students have to know how to verify.

How we do it

This is not theory. At Dixi-Project we work this way every day: AI writes the code, no agent can touch approved tests, a person signs every commit and states that they answer for the change, and an automated Gate verifies before anything reaches production. AI does not sign, approve or merge. We describe it in One maintainer, one agent and 141 PRs with no human review and in Inside our spec-driven development framework.

What needs to be taught is how to take the place of the person who signs.

What could go wrong

  1. Accreditation becoming a formality. If the evidence is requested on paper and nobody reviews a real audit, proposal 1 changes nothing.
  2. Scale. Oral defenses work in groups of 30, not 120. That is why the seeded-defect audit, which grades itself, is central.
  3. Vendor lock-in. An agreement with a single AI company solves compute this year and creates a problem the next. Hence open models.
  4. Industry not creating apprentice roles. Without incentives, proposal 9 depends on goodwill, and goodwill does not scale.

What this article does not claim

  • That these ten proposals are validated. They are a reasoned design, not tested policies.
  • That we speak for CACEI, TecNM, ANUIES, SECIHTI or any other institution we mention.
  • That we know what AI will be able to do in five years. The proposals bet on what holds in any scenario: fundamentals, verification, domain knowledge and accountability.

Sources

  • Tecnológico Nacional de México, Computer Systems Engineering curriculum (2010 plan).
  • Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine?, Stanford Digital Economy Lab, August 2026 version.
  • Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS 122(26) (2025).
  • Kwa et al., Measuring AI Ability to Complete Long Software Tasks, METR (2025).
  • INEGI, Censos Económicos 2019.
  • Ley Federal de Protección de Datos Personales en Posesión de los Particulares, Diario Oficial de la Federación (March 20, 2025).
  • Dixi-Project: What to study (and what not to) in a systems engineering degree for 2032, The broken pipeline, Verification debt and Who answers when AI wrote the code? (2026).
ShareLinkedInXWhatsApp