From "Coding Sweat-shops" to "Trusted Agent Factories": The Resilience of the Peripheries in the Age of Augmented Intelligence

One of those posts written in the small hours by an anonymous coding geek, probably somewhere on the Costa del Sol or in Getafe, in a moment of anger and inspiration, has set off the article that follows.

The reflection that triggers this analysis —the normalisation of a 30% or 40% pay gap for southern European engineers compared with their northern counterparts, under the pretext of "cost of living"— is not a simple workplace complaint found on LinkedIn, it is the local symptom of a global tectonic fault. Because it is worth saying from the outset: what an engineer in Málaga or Lisbon experiences as a comparative grievance, an engineer in Bangalore, Manila or Ho Chi Minh has been living for three decades as a business model. The coding sweat-shop was not born in the Mediterranean; it was born in Asia, it was perfected in Asia, and today it is threatening to go extinct —also in Asia— under the weight of Agentic Artificial Intelligence.

Taking on the role of an economist, we can observe in this anonymous complaint the tearing apart of "money illusion" and the deep corporate confusion between labour intensity (human effort) and marginal productivity (added value). But the diagnosis of this imbalance is only the prologue. We are standing at the threshold of a violent transition towards Agentic AI, a paradigm shift that not only threatens to destroy the current model of outsourced technological execution, but demands a radical reconfiguration of talent across every periphery of the system: from the Guadalquivir to the Ganges, from the Turia to the Mekong.

What follows is an essay that dissects the structural problem of technological monopsonies, the upheaval brought about by agentic AI, and a solution that, this time, deliberately resists simplification: two levers are not enough; what is needed is a complete ecology of factors.

We are standing at the threshold of a violent transition towards Agentic AI, a paradigm shift that not only threatens to destroy the current model of outsourced technological execution

1. The Problem: The Illusion of Productivity and the Global Geography of Arbitrage

The contemporary tech labour market has been structured around a geography of inequality that splits the world into two mutually dependent but economically asymmetric categories: the centres of decision and the centres of execution.

1.1. The Cost-of-Living Fallacy and the Asymmetry of Value

For years we have normalised the idea that a developer in Madrid, Barcelona, Málaga or Lisbon should earn substantially less than someone doing exactly the same job in Munich, London, Paris or Dublin. The corporate defence has always been the adjustment to local cost of living. Yet this argument collapses in the face of the evidence of price globalisation, all the more so with the proliferation of digital nomads.

If you work in tech, you are anchored to urban ecosystems where non-tradable goods (fundamentally, housing) have undergone ferocious inflation, converging with northern European prices. At the same time, tradable goods (vehicles, electronic devices, software subscriptions) operate under the global Law of One Price. The "real saving" that justifies the pay gap is, mathematically, a mirage. And the mirage repeats itself, with even more brutal multipliers, in Bangalore —where housing prices have risen faster than in many European capitals— or in Mexico City.

1.2. The Asian Origin of the Model: from Body Shopping to the Pyramid

The term coding sweat-shop has a genealogy. It is born out of Indian body shopping in the nineties: the export of engineers in batches, billed by the hour, to projects in the global North. On that logic the pyramid model of the big IT services firms was built —Tata Consultancy Services, Infosys, Wipro, HCL, Cognizant—: mass recruitment of junior engineers at the base, billing per FTE (full-time equivalent), and margins built on the difference between what the parent company charges the Western client and what it pays the developer in Hyderabad or Pune. A sector that came to employ millions of professionals and to sustain a significant share of India's GDP on a single premise: selling cheap human time.

The model has regional variants worth naming. In China, the 996 culture —nine in the morning to nine at night, six days a week— pushed labour intensity to extremes that the European vocabulary of burnout cannot even begin to describe. In the Philippines, Manila and Cebu built on top of BPO a growing layer of development and QA billed by the hour. Vietnam, Bangladesh and much of Eastern Europe (Cluj, Wrocław, Belgrade) completed the map of peripheral execution.

And there is an even deeper basement in this architecture: the data annotation factories. In Nairobi, Manila or Caracas, thousands of workers label images, moderate toxic content and fine-tune, through human feedback, the very language models that today threaten to replace programmers. The irony is complete: the most precarious sweat-shops on the planet are the ones that have fed the technology that will devour the code sweat-shops. It is the Synthetic Feedback Loop in its cruellest version.

The irony is complete: the most precarious sweat-shops on the planet are the ones that have fed the technology that will devour the code sweat-shops.

1.3. The Economic Diagnosis: Oligopolistic Monopsony and Digital Vassalage

The real engine behind this gap is not the cost of living, but a market structure operating under a regime of oligopolistic monopsony. Large corporations have set up their strategic headquarters in the northern hubs (Silicon Valley, London, Munich, Dublin), where they concentrate decision-making, intellectual property, financial engineering and the design of the business architecture. The peripheral geographies —Málaga, Lisbon, Medellín, but also, and above all, Bangalore, Manila or Ho Chi Minh— have been systematically categorised as software factories billed by the hour.

In these centres, the business model is not based on generating intellectual property, but on labour arbitrage. The goal is to maximise the output of lines of code and the resolution of tickets at the lowest possible cost.

Something China knows very well and has ended up correcting in its 15th five-year plan (2026-2030) by promoting technological and data self-sufficiency.

1.4. The Confusion between Effort and Productivity

Here lies the perversion of the system: the confusion between human effort and economic productivity. In the coding sweat-shops —whether under the Andalusian "sunshine tax" or under Shenzhen's 996— we see teams stretched to the limit, professionals taking on multiple roles and exhausting working days. There is absolute attrition. And yet, from a macroeconomic perspective, productivity is defined as the economic value generated per unit of time. If the parent company in London retains the rights over the product and sells the licence at a global price, while the centre in Málaga —or in Chennai— only bills "development hours" at a reduced internal rate, that developer's economic productivity looks artificially low. Not because they work little, but because the corporate structure prevents them from capturing the real value of what they produce.

This is a model of digital vassalage in which the periphery takes on the heavy operational burden and the imperial centre absorbs the margins, justifying it with a false metric of productivity.

If the parent company [northern tech hub] retains the rights over the product and sells the licence at a global price, while the centre [southern coding sweat-shop] (...) only bills "development hours" at a reduced internal rate, that developer's economic productivity looks artificially low. (...) a model of digital vassalage in which the periphery takes on the heavy operational burden and the imperial centre absorbs the margins, justifying it with a false metric of productivity.

2. The Disruptive Impact of Agentic AI: A Mass Extinction Event

If the model described was unfair but stable, the arrival of Agentic Artificial Intelligence (autonomous systems, Agent-to-Agent or A2A architectures) acts as a mass extinction event for this status quo.

2.1. The Collapse of Transactional Execution

Until now, Generative AI acted as a copilot: it helped the developer in Barcelona, Medellín or Pune write code faster. This increased the intensity of the factory, but kept the human at the centre of execution.

Agentic AI changes the equation at its root. Autonomous agents don't just suggest code; they plan architectures, run continuous testing, talk to other agents to fix errors, deploy to production and monitor results. The basic software development lifecycle —exactly the product that coding sweat-shops sell by the hour— gets automated.

When we audit multi-agent architectures inside large financial or retail organisations, we observe that the transition towards autonomous delegation is not happening calmly or with long-term planning; on the contrary, adoption is being driven by urgency, anxiety and a sense of pressure. The pressure from financial markets to maximise margins is pushing parent companies to replace much of their peripheral outsourcing contracts with swarms of autonomous agents.

Autonomous agents don't just suggest code; they plan architectures, run continuous testing, talk to other agents to fix errors, deploy to production and monitor results. The basic software development lifecycle —exactly the product that coding sweat-shops sell by the hour— gets automated

2.2. The Inverted Pyramid: the Asian Earthquake

Where this collapse takes on seismic dimensions is in Asia. The Indian pyramid model depends structurally on its base: tens of thousands of junior engineers hired every year for maintenance, migration and support tasks. It is precisely that base —transactional code, manual QA, ticket resolution— that is the first layer autonomous agents swallow. A pyramid without a base is not a pyramid: it is an obelisk about to topple over. The Indian IT services firms themselves know it, and their recent moves —freezing mass hiring of freshers, a rhetorical pivot towards "platforms" and "outcomes" instead of FTEs— are the implicit confession that the model of billing per human body has entered its terminal phase.

The same diagnosis applies, with nuances, to Manila, to Ho Chi Minh and to the nearshore factories of Eastern Europe. An AI agent deployed in the cloud does not understand geographies. Its marginal cost of execution tends to zero and is identical whether it runs from a server in Frankfurt, in Madrid or in Singapore.

Image generated by OpenAI

2.3. The Destruction of Labour Arbitrage

If the only competitive advantage of a hub in Lisbon, Medellín or Bangalore is offering talent at a 40% —or an 80%— discount compared with Dublin or San Francisco, that hub is doomed to irrelevance in the short term. Agentic AI destroys the "billing by the hour" business model of pure transactional execution.

This scenario creates a profound risk of cognitive offloading and mass job displacement. If the geographies that are intensive in code factories do not react, we will witness the economic decay of these poles, turning them into digital industrial deserts. And it is worth underlining the scale: if in southern Europe we are talking about tens of thousands of jobs, on the Indian subcontinent we are talking about millions.

2.4. The Exception That Lights the Way: India's GCCs

And yet, Asia also offers proof that mutation is possible. Against the model of outsourcing by the hour, India has seen the flourishing of Global Capability Centres (GCCs): multinationals' own centres that have evolved from cost arbitrage to ownership of capabilities —product engineering, data science, risk governance, even global architecture decisions—. The GCC that designs a pharmaceutical company's data strategy from Hyderabad is no longer a sweat-shop: it is a relocated centre of decision. The lesson is clear: geography does not condemn you; your positioning in the value chain does.

3. The Solution: The Pivot towards "Trusted Agent" Factories

Faced with the imminent annihilation of pure execution, the technological peripheries —of southern Europe, of Latin America and of Asia— do not have the option of resisting passively; they must mutate aggressively towards a higher stage in the value chain: becoming Governance Laboratories and Trusted Agent Factories.

At first glance, this migration seems to pivot on two pillars: the deep development of critical thinking and certified expertise in agentic governance. It is a pedagogically useful reduction —and I will keep it as scaffolding—, but it would be intellectually dishonest to present it as sufficient. No territory has ever climbed a value chain with two levers. What is needed is a complete ecology of factors, and I will devote the second half of this section to it. First, the formula that governs the whole argument:

Trust is the exponent that lifts technology up or sinks it. And trust is not programmed; it is audited, governed and reflected upon.

Pillar A: The Governance of Agentic Artificial Intelligence

For a centre in Málaga, Medellín or Pune to stop being a supplier of cheap hours and become an indispensable strategic partner, it must specialise in what machines cannot regulate on their own: security, corporate ethics, regulatory compliance and the mitigation of systemic risk.

When a parent company in Munich or Silicon Valley deploys multi-agent architectures to make financial or investment decisions, the main risk will not be the cost of the code, but "cascading hallucination", algorithmic bias and regulatory non-compliance (such as the European AI Act). This is where the old code factories must deliver incalculable value by taking on the role of auditors and guarantors of trust. Resilience means mastering —and being able to certify— three emerging disciplines:

1. The systemic management of artificial intelligence. Just as quality or information security ended up being governed through formal management systems, AI is following the same path: there is already an international standard that requires designing, implementing and auditing complete AI management frameworks within the organisation (e.g. ISO/IEC 42001). A hub in Barcelona —or in Bangalore— that masters this discipline stops selling code and starts selling Operational Certainty: the guarantee that a corporation's swarms of autonomous agents operate responsibly, transparently and traceably, with algorithmic risk embedded into traditional corporate governance.

2. The cognitive audit of machines. Deep technical auditing is the new QA. While agents write the transactional code, we will need professionals capable of auditing the hidden logic of models: the quality of training datasets, the explainability of agentic decisions, compliance with ethical and regulatory frameworks. It is the transition from software testing to cognitive auditing, a practice that the major professional associations of systems auditing have already begun to codify and accredit.

3. The security of agentic systems. Autonomous agents interacting with one another (A2A) open up unprecedented attack vectors: indirect prompt injection, agent hijacking, manipulation of system memory. Peripheral hubs must become fortresses of agentic cybersecurity: cryptographically sealed autonomous communication, emergency shutdown protocols, damage containment tested and validated. Here too, specialised training paths in AI security management are starting to consolidate.

The specific certifications —they already exist, both in the international standards arena and in that of the professional auditing and security associations (e.g. ISO/IEC 42001, ISACA)— matter less than the capability they certify. What is decisive is not the seal, but the verifiable mastery of these three disciplines. In moving towards them, the peripheral hub takes ownership of the most critical link in the value chain: the mitigation of catastrophic risk. Parent companies will pay a very high premium for teams that guarantee their autonomous agents do not destroy their reputation or their balance sheet in milliseconds.

Pillar B: Critical Thinking Squared (CT²)

Technical governance and certifications are not enough if the people applying them are not capable of interpreting ambiguity, detecting biases and making decisions under uncertainty. In the agentic era, the human advantage will not lie in executing instructions better than a machine, but in formulating better questions than a machine.

That is why the second pillar is not simply critical thinking, but Critical Thinking Squared (CT²): the ability to examine not only the problem, but also the mental frame from which we are trying to solve it. Put another way: auditing the agent is not enough. You have to audit the auditor too.

This distinction will be decisive. Autonomous agents will be able to write code, generate hypotheses, run tests, optimise processes and propose decisions at a speed unattainable for any human team. But precisely for that reason we will need professionals capable of stopping in front of an apparently flawless recommendation and asking, in an augmented way:

• what invisible assumption holds this conclusion up?

• what data is missing?

• what incentive is distorting the result?

• what harm could be caused by a decision that is technically correct but humanly unjust?

That will be the new differentiator for tech hubs that want to escape the cheap-execution model. Málaga, Lisbon, Medellín, Bangalore or Manila will not be able to compete against agents writing transactional code. But they will be able to compete by building teams capable of governing synthetic reasoning with human judgement, synthetic collaboration, ethical depth and intellectual discipline.

Critical thinking in this context cannot be reduced to logic, fallacy detection or evidence analysis. All of that matters, but it is not enough. A system can reason impeccably and still produce an absurd, unjust or socially harmful decision. History is full of technically correct syllogisms in the service of moral atrocities.

That is why CT² demands at least four integrated capabilities.

First, intellectual humility: accepting that our frame may be incomplete, biased or plainly wrong. In agentic governance, this humility is not a decorative virtue; it is a safety mechanism.

Second, metacognition: observing how we think while we think. Asking ourselves whether we are evaluating the evidence or defending a prior preference dressed up as rigour.

Third, contextual judgement: understanding that an algorithmic decision never happens in a vacuum. It always has an impact on people, institutions, incentives, cultures and power relations.

Fourth, social contrast: subjecting our reasoning to other perspectives. Critical thinking does not mature in isolation. It needs friction, peer review, cognitive diversity and uncomfortable conversation.

This last dimension is especially important. Trust is not born of certified individuals working in silence in front of a screen. It is born of professional communities capable of arguing, correcting themselves and raising the collective standard of judgement.

Here the technological peripheries have a real opportunity. Their cultural, linguistic and economic diversity can become a strategic advantage if it stops being treated as cheap labour and starts operating as collective intelligence. The Global South must not limit itself to executing code designed from the centres of technological power; it must take part in defining the criteria by which those systems will be audited, governed and legitimised.

Because governing autonomous agents is not only a technical problem. It is an institutional responsibility. And, ultimately, a human one.

A hub that combines certifiable governance with CT² stops selling hours. It starts selling something far scarcer: trustworthy judgement under uncertainty.

That is the mutation. From the programmer as executor to the professional as guarantor. From the team that resolves tickets to the team that protects decisions. From the code factory to the trust factory.

Beyond the Two Pillars: The Ecology of Factors of the Transition

Is it honest to claim that critical thinking plus governance certification "solves everything"? No. They are necessary conditions, not sufficient ones. No peripheral territory will climb the value chain without an ecosystem of complementary factors interacting with one another.

I propose at least six additional vectors:

1. Sovereignty over intellectual property. The sweat-shop sells hours; the Trusted Agent factory must sell assets. That requires peripheral hubs to generate and retain their own IP: vertical agents packaged as product, proprietary audit frameworks, licensable methodologies. Without patient capital —venture, corporate or public— to finance the shift from services to product, and without ownership structures that let talent capture the value it creates, certification only produces better-trained vassals.

2. The regulatory advantage of proximity. Here southern Europe holds an asset that Bangalore does not have: it lives inside the perimeter of the AI Act. Regulatory proximity —linguistic, legal, cultural— to the European legislator turns Lisbon, Barcelona or Madrid into natural interpreters of the most demanding regulatory framework on the planet, exportable via the "Brussels effect". Latin America can play the symmetrical card as a regulatory bridge between legal traditions; Asia, the card of scale and deployment speed. Each periphery must read its own regulatory geography as a competitive advantage, not as a burden.

3. Physical and energy infrastructure. Agents do not live in the ether: they live in data centres that consume energy. Southern Europe —with its abundance of solar and wind— and several Latin American geographies have the opportunity to physically anchor the agentic economy in their territory: sovereign compute, local latency, cheap renewable energy. Whoever hosts the compute negotiates from a very different position than whoever merely rents out brains.

4. Cultural intermediation. There is an irreducibly human differential that no agent replicates: the ability to translate between corporate cultures, to read what a board of directors does not say, to spot that a technical requirement conceals a political or cultural conflict. Multilingual and multicultural peripheries (Spanish as a bridge between Europe and the Americas, Asian technical English, Filipino biculturalism) hold here a strategic raw material that must be deliberately cultivated, not taken for granted.

5. Institutional alliances for reconversion. The mutation of tens of thousands —or millions— of executors into auditors and governors of agents will not happen by spontaneous generation. It demands pacts between universities, business schools, public administrations, multilateral bodies and corporations to reconvert talent at industrial scale and speed. India understood this with its national missions of reskilling.

6. Community density of the ecosystem. The social contrast I spoke of in CT² has its territorial counterpart. The hubs that survive are not collections of certified individuals, but dense communities where knowledge circulates, is contrasted and recombined —meetups, professional chapters, governance forums, permanent intellectual friction—. Trust, once again, is a community phenomenon before it is an individual one.

These six vectors are not a menu to choose from: they are a system. IP without critical talent is sterile. Talent without infrastructure emigrates, and regulation without community becomes bureaucratised. And the whole set —pillars and vectors— is still raised to the same exponent: Trust.

IP without critical talent is sterile. Talent without infrastructure emigrates, and regulation without community becomes bureaucratised. And the whole set —pillars and vectors— is still raised to the same exponent: Trust.

Synthesis: Resilience Lies in Community and Dignity

The destiny of peripheral regions is not written in stone. The transition from coding sweat-shops —subjected to monopsony and labour arbitrage, from Málaga to Manila— towards genuine Trusted Agent Factories is the greatest industrial challenge of the coming decade. The GCCs (Global Capability Centres) in India prove that it is possible; the annotation factories of Nairobi remind us what happens when no one even tries. It is perhaps the third way in the AI economy.

Getting there takes more than squeezing models or teaching new programming techniques. It takes more than —let us be rigorous with ourselves— critical thinking and governance certificates. What is required is cultivating a complete ecology: intellectual, ethical, institutional, energy-related and communal. The regions of Málaga, Barcelona, Lisbon, Medellín, Bangalore and Manila must bet on hybridising the strictest corporate governance with a deep humanism, and on the patient construction of the assets —intellectual property, infrastructure, community— that turn certification into sovereignty. This bet will transform the spirit of resilience into financially robust proof of the "wisdom of the crowd", showing how the shared judgement of a heterogeneous and autonomous collective outperforms, in terms of value judgement, the isolated perspective of any individual specialist. This is the thesis of resilience transformed into value by the ecology of collaboration.

Developing critical thinking in the age of augmented intelligence requires accepting the responsibility of exposing oneself to the evidence and listening to other minds with a genuine desire to learn. And developing it squared requires, on top of that, the humility to audit our own formulas —this one included—. Critical thinking is not just an individual cognitive skill: it is a form of intellectual responsibility exercised in community.

I learned this in the simplest of ways, in a conversation with my partner Gam Dias , while writing The Data Mindset Playbook. I had just come back from Saudi Arabia and shared with him a reflection about the injustice I felt at polygamy still being legal. My argument was, deep down, intuitive: I saw an obvious inequality between men and women. Gam, critical by nature and of Sri Lankan origin, answered me with an uncomfortably illuminating question: “would it seem less unjust to you if a woman chose several men?”.

It was not a provocation. It was a gentle invitation to look at the blind spot in my own cultural frame. Polyandry —a woman married to several men— exists in various South Asian traditions, among them parts of India, Sri Lanka or Nepal. The anecdote did not resolve the moral dilemma, but it did reveal something more important for this piece of writing: even our ethical intuitions need cultural contrast. Diversity does not always give us immediate answers, but it almost always improves the quality of our questions.

And this is precisely the lesson that the technological peripheries must bring to the terrain of agentic AI. If we aspire to design systems where machines act on behalf of humans, we cannot govern them from a single cultural, economic or geopolitical grammar. We need the code centres of the Global South to stop being merely spaces of execution and to become spaces of reflection, auditing and legitimacy through the direct action of diversity.

If agentic AI aspires to become the universal fabric of every industry, its governance cannot be born of a narrow, bipolar and geopolitically captured vision. It needs a layer of reflection that is truly universal, diverse and eminently human. A layer capable of transcending the current map, where the United States and China concentrate the centres of technological decision while the rest of the world is relegated to the role of "AI laggard" or peripheral coding sweat-shop.

Ultimately, it is extremely difficult, not to say impossible, to govern autonomous agents without a vision oriented to the common good and without a non-negotiable perspective on human dignity. The regions that understand that the future of technological work does not lie in mechanised execution —whether they are in Andalusia, Antioquia or Karnataka—, but in ethical auditing, the governance of trust and critical rigour raised to the square, will not only survive agentic AI: they will lead the global knowledge economy, finally breaking the chains of North-South global technological vassalage.

This is an open hypothesis and I hope it is read as such. A first trigger for an exercise in reflective ecology.