You added a coding agent to your workflow. You’re faster now. Features that used to take a day take an afternoon. It feels like a superpower, and for a while it is.

But look down for a second. You’re still in the traces. You’re pulling the cart right alongside the machine, hand on the keyboard, correcting it, re-prompting it, cleaning up after it. “AI-assisted” means a human and an agent hauling the same load together. That’s better than one human hauling it alone. It is not the thing.

Here’s the uncomfortable truth: as long as a human is in the loop of producing the work, you’ve only made the human a faster laborer. And a faster laborer is still spending the two things you can never buy back: their time and their attention. Every hour someone on my team spends generating output by hand is an hour they didn’t spend on the strategy, the IP, the judgment that no agent has. That’s the wrong trade, and doing it faster doesn’t fix it. It just breaks even sooner.

So we drew a line. At HikmaAI, we don’t code manually. Not the app logic, not the tests, not the plumbing that runs the repo itself. Humans don’t hand-patch. That sounds extreme written down, and it felt extreme the first time I said it out loud, coming from someone who has spent most of his adult life happy with his hands in an editor. But it’s not a stunt. It’s a direction.

Think of a harness

A workhorse is strong. Left loose in a field it’s still strong, and completely useless to you: the power has nowhere to go. You don’t make the horse stronger by grabbing a rope and pulling next to it. You make it useful by putting it in a harness.

A good harness doesn’t add horsepower. It channels the power the animal already has. And it does one more thing that matters more than people think: it moves the human out of the traces and up onto the seat. You stop pulling. You start driving. Reins, not rope. You point, you correct, you decide where the whole thing is going, and the power under you does the hauling.

That’s the shift. The agent is the horse. It has real power now, more every month. What’s been missing isn’t strength. It’s the harness.

So we built one. We call it Bridles. I’m not going to walk you through its guts today, but the idea behind it is simple enough to say in a paragraph: it turns our engineering conventions into something an agent can actually read and follow, and it refuses to call work “done” on the agent’s word alone. Humans plan, steer, and validate. The agent does the rest.

Three moves, and only three, are ours.

The three things a driver actually does

Plan. Point the cart at the right place before anything moves. This is where taste lives, where the hard product calls get made. It’s the most human part of the job and, not by accident, the part I want my people spending their hours on.

Steer. Not constant correction. A hand on the reins for the moments that genuinely fork, where the road splits and only judgment tells you which way. The rest of the time, you let it run.

Validate. Not “does it look right.” Does the evidence agree that it is right. This is the part most people skip, and skipping it is how you end up trusting a confident machine that’s confidently wrong.

Everything else that used to eat a developer’s day now happens down in the harness, out of the human’s hands.

What I learned building it

A couple of things surprised me, and they’re worth more than the tooling.

The first is about being stuck. When an agent gets stuck, the old instinct kicks in immediately: try harder, write a cleverer prompt, or just grab the keyboard and do it yourself. We banned that instinct. When the agent gets stuck, that’s not a signal to jump into the traces. It’s a signal that something is missing from the harness: a guide it can’t see, a check it doesn’t have, a piece of context that lives in someone’s head instead of in the repo. So you don’t do the work by hand. You build the missing piece, and then you let the horse pull again. Struggle stopped being a reason to intervene and became a to-do list for the harness.

The second is about trust. I don’t trust the agent because it sounds confident. Confidence is free, and a model has an endless supply of it, and it will spend every bit of that confidence being wrong. So the whole system runs on evidence instead of on prose: nothing is “done” because the agent said so. That one rule, evidence over eloquence, changed how the team works more than any model upgrade did.

From context engineer to harness engineer

Here’s the reframe, and it’s the real point of this post.

For the last couple of years the craft was context engineering: feeding the model the right information, at the right time, in the right shape, so it would do a good job on the next task. Useful. Still necessary. But it keeps you next to the horse, handing it things.

The craft now is harness engineering. The question is no longer “how do I get the agent to write this code.” It’s “how do I build the thing that lets the agent write this code, and reliably prove it’s right, without me in the traces.” That is a different job. It requires:

• Encoding what your best people know into guides an agent can follow, instead of keeping it in their heads
• Building the checks that decide “done,” so a human doesn’t have to read every line to trust it
• Treating every stuck moment as a missing capability to build, not a task to do by hand
• Getting comfortable up on the seat, steering, when every instinct trained over a career says grab the reins and pull

I’ll be honest: that last one is the hardest, and I’m not fully past it. The pull to open the editor and “just fix it myself” is strong. Sitting on the seat and trusting the harness is a discipline, not a feeling. I’m still building the muscle.

I’m going to write about this transition as we live it. What breaks, what the team resisted, where the harness wasn’t good enough and we had to go rebuild it. Not a polished case study after the fact. The real thing, as it happens.

Ask the machine first

Here’s the habit I keep pushing on my team, and it’s smaller than the harness but it feeds it. Before you start any task, before you open the doc or the editor or the spreadsheet, ask one question: can the agent do this instead of me? Not “can the agent help me do this.” Can it do it.

That question sounds trivial. It isn’t. It rewires everything upstream and downstream. It changes what you hire for, because you stop staffing the work and start staffing the judgment. It changes execution, because your first move on any task is no longer “how do I do this” but “how do I hand this off and check it.” Make it your reflex and the harness stops being a thing engineering built. It becomes how the whole company thinks.

And once you’re asking it, you notice how much work you’ve been quietly not doing. The stuff that never makes it off the someday list because you don’t have the hours. Scraping arxiv.org for the papers that actually touch your business. Pulling the YouTube channels you follow, transcribing the talks worth an hour of your attention, and getting back the three paragraphs that matter instead of the ninety minutes. Watching Reddit so you don’t have to. The point isn’t to read more. It’s to triage: let the agent do the reading and hand you the shortlist. Same move on the rest of it, the issues piling up in your codebase, the leads nobody’s chased, the marketing idea you had in the shower and lost by lunch. That’s not laziness dressed up. It’s the whole bet. Work smart, not hard.

This isn’t an engineering trick

The last thing, and the reason I think this matters beyond my four walls.

None of what I described is specific to code. Guides that carry your best judgment. Sensors that decide when something is actually done. Humans reserved for planning, steering, and validating, and pulled out of the manual grind. That’s not an engineering pattern. That’s an operating model.

The same harness thinking belongs in marketing, in sales, in every function that today runs on people hand-producing work an agent could pull. Engineering is just where we started, because it’s where the sensors are cheapest to build and the feedback is fastest. It is not where it ends.

The goal was never a company of faster laborers. It’s a company where every team is up on the seat, hands on the reins, driving. The horses are strong enough now. The work is building the harness, and then having the nerve to let go of the rope.

— Mauro

References:

• Martin Fowler on harness engineering (guides and sensors)
• OpenAI on exec-plans and zero-hand-written-code harness engineering
• Alexis King, “Parse, don’t validate”
• LangChain on agent-cost governance

On long-horizon agentic benchmarks, open-weight models with the right harness and task-specific tuning are posting 20+ point gains over frontier baselines. Not a marginal edge. Twenty-plus points, on the work that actually runs in production: agents calling tools, orchestrating workflows, operating autonomously for hours.

The number is concrete, not hand-waved. WebRL, a curriculum reinforcement learning framework out of Tsinghua and Zhipu, took the open GLM-4-9B from 6.1% to 43.0% on WebArena-Lite, clearing GPT-4-Turbo at 17.6% and GPT-4o at 13.9%. Same benchmark. A model a fraction of the size. And it wins.

Where frontier models still win (and why that isn’t the point)

Frontier models are still superior for narrow, high-complexity reasoning. Legal analysis. Deep research. Specialist inference. The hard, single-shot problems where one brilliant answer is worth almost any price. Nobody serious is disputing that.

Paolino Madotto nailed this in Agenda Digitale: the value of AI is shifting from the model itself to the system that orchestrates it. In a well-designed system, the frontier model is not the engine running every token of an hour-long agent loop. It’s the premium specialist the orchestrator calls, selectively, when a problem genuinely demands it.

That’s frontier’s rightful place: the expensive expert on retainer, not the machine you leave running all day.

And notice where the leverage actually sits. Harness engineering alone (the loop, the tools, the retry logic, the verification wrapped around a fixed model) can buy 10-15 points on some agentic benchmarks without touching the weights. The scaffold is doing the work. Which is exactly why the model you drop into that scaffold can be smaller, cheaper, and yours.

The metric almost nobody is measuring

Everyone quotes cost per token. Almost nobody making the actual buy-or-build decision measures watts per token: energy per token. That’s the number I keep coming back to, because it’s the one that decides whether an agentic system survives contact with scale.

The researchers are starting to catch up. Stanford’s Hazy Research group calls it intelligence per watt, and their numbers point the same way: IPW improved 5.3x in two years (3.1x from better small models, 1.7x from hardware), and local models already handle 88.7% of single-turn chat and reasoning queries. The trend line isn’t subtle.

Here’s the math that matters.

A long-running agent doesn’t emit a few thousand tokens. It loops. It retries. It calls tools and reasons over the results for hours. It burns millions of tokens per task, per run, per user. Multiply watts per token by that volume, across a fleet of agents in production, and energy stops being a footnote on the invoice. It becomes the business case. Or the reason the business case collapses.

Frontier APIs were priced for the era of the single clever answer. Agentic workloads live in a different regime entirely: enormous token volume, most of it routine orchestration that doesn’t need a frontier brain. Paying frontier rates for that volume, in dollars and in watts, isn’t a strategy. It’s a leak.

Why the token economics don’t hold

Current token economics break at agentic scale. Not because API prices are high today, but because prices and physics obey different laws. Prices can drop: they’re a margin decision. Joules per token have a floor, and that floor is set by model size. Every token a trillion-parameter model generates costs far more energy than the same token from a 9B model, no matter whose datacenter runs it.

And let’s be honest about the fine print: the watts-per-token win doesn’t come from the server’s zip code. A local GPU idling at 30% utilization can burn more energy per token than a hyperscaler’s batched fleet. The win comes from right-sizing the model to the task and keeping the hardware busy. What local adds is the thing no API can sell you: complete data sovereignty, not a single byte leaving the building.

Right model. Right task. Right infrastructure.

That’s not a compromise. That’s good engineering.

Where this leaves security (and where HikmaAI comes in)

Owning the full stack means owning the full stack. That includes the security layer. Here’s the part too many “go local” arguments skip.

The moment you bring models in-house and let agents operate autonomously, you bring the entire security and governance burden in-house with them. A frontier API at least ships with the provider’s guardrails baked in. An open-weight model you run yourself has none of that by default. And an agent that calls tools and acts for hours is a far larger attack surface than a single chat completion. Prompt injection, goal hijacking, PII leakage, tool misuse: these scale with autonomy. That’s the hidden tax on the watts you just saved.

Going local doesn’t shrink the need for a security platform. It multiplies it. That’s the problem we built HikmaAI for.

HikmaAI is platform-agnostic by design. It scans and evaluates your models and your agents (Ollama, OpenAI, or your own REST API; proprietary or open-weight; cloud, on-prem, or hybrid) for security vulnerabilities, prompt injection, toxicity, PII exposure, and compliance gaps, with adaptive red teaming and guardrails for the systems you now own end to end.

And it does it without breaking the promise that made you go local in the first place: your data stays on European infrastructure, no source code leaves your environment, nothing crosses a border it shouldn’t. The same sovereignty argument that drives the move to local models is the argument for governing them with a platform built around sovereignty from the ground up.

Small, local, efficient, and secured. That’s the full stack.

Right model. Right task. Right infrastructure. Right guardrails.

— Mauro

References:

• Paolino Madotto, “Il vero valore dell’AI non è più nel modello, ma nel sistema che lo guida,” Agenda Digitale.
• Qi et al., “WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning,” arXiv:2411.02337.
• Saad-Falcon, Narayan et al., “Intelligence Per Watt: A Study of Local Intelligence Efficiency,” Stanford Hazy Research, blog / arXiv:2511.07885.

The US turned off its best AI for the rest of the world in the time it takes to sign a letter. Europe should treat that as the most useful thing that has happened to it in years.

Let me say plainly what yesterday meant, because the polite version helps no one.

On 12 June the US government ordered Anthropic to shut off its two most advanced models, Fable 5 and Mythos 5, for every foreign national on Earth. Not just for people abroad. For foreigners living and working inside the United States too, including Anthropic’s own engineers, locked out of the systems they helped build. The order did not come from a safety regulator. It came through export-control authority, the same machinery that governs the shipment of weapons and advanced chips. Anthropic could not sort its users by passport, so it did the only thing it could. It turned the models off for everyone, three days after launching them. (Anthropic’s statement; Al Jazeera)

Read that timeline again. A model that hundreds of millions of people had already started to depend on was switched off, worldwide, by a single letter, in seventy-two hours. No vote. No warning. No appeal.

That is not a story about one company, or one administration. It is a story about where Europe actually stands. So let me think it through the way I have been turning it over since yesterday, point by point.

One. The frontier is American, and now it is policy.

US companies build the most capable models. The rest of us get them later: or, as of yesterday, not at all. We told ourselves that gap was a few months of market timing, an irritation we could plan around. It was always more than that, and now it is explicit. Access to the best intelligence on the planet is a lever, and the US has shown it is willing to pull it. The delay was never just a by-product of where the labs happen to sit. It is an instrument, and we are on the wrong end of it.

Two. Exclusive access compounds, and it compounds against us.

A model that is both far stronger and available to only one side does not help that side a little. It helps it everywhere, every day, on every problem its companies touch. The American firm gets the better tool and we get the announcement. Stack that edge across a few product cycles and you no longer have a gap you can close with effort. You have European companies, whole industries, quietly outrun by competitors whose only real advantage was the better machine and the permission to keep it. We could lose entire sectors, not in a fight we entered and lost, but in one we were never allowed to join.

Three. And even if we win, the hardware closes the loop.

Say we do the hard thing. We build the datasets, we train the models, we get genuinely good. Those models still run on chips we do not make. Nvidia designs roughly 90% of the world’s AI silicon and will not sell us its best, because the advanced parts are rationed by the same export rules that just silenced Anthropic. So we would own the data and the models and still be renting the metal underneath them: straight back to points one and two, one layer down. Who really owns what runs on someone else’s machine?

Now the part that is hard to say, because saying it makes the problem enormous.

You cannot fix this one layer at a time. “Data sovereignty”, the comfortable phrase everyone in Brussels reaches for, is not sovereignty. It is a label we stick on the act of integrating someone else’s systems and hoping they stay switched on. If the models can be revoked, the chips withheld, the cloud compelled, then owning the data is owning the one layer that matters least. Sovereignty is not a layer you can buy. It is the whole stack, or it is theatre. Chips, models, data, infrastructure. All of it, or none of it counts.

I know how that sounds. Slow. Expensive. Close to impossible. We are talking about a project measured in a decade, not a budget cycle, and Europe is not famous for acting as one country for ten years at a stretch. But here is what I keep returning to. We are not starting from zero, and we insist on behaving as if we are.

The single most irreplaceable machine in the entire AI supply chain is not American. It is Dutch. Every advanced chip on Earth, including every Nvidia part the US is currently rationing, is printed by a lithography machine that exactly one company in the world knows how to build. That company is ASML, in Veldhoven. The Americans cannot make the best chips without us either. We are not a bystander in this supply chain. We own its narrowest chokepoint; and we have never once used it as leverage, because we have trained ourselves to be the customer in every layer instead of the owner of the one layer that cannot be replaced.

And it is not only the chokepoint. A few streets from the same Dutch supply chain, in Eindhoven, a company called Axelera AI is designing its own AI silicon: its Europa processor launched in 2025, it has raised more than 450 million dollars, and its next chip rides on Europe’s own RISC-V sovereignty programme. One company does not make a continent independent. But it is proof, sitting right there, that the talent and the design know-how are not a thing we have to invent from scratch. We are not starting from zero. We keep pretending we are.

So here is what the wake-up call is really asking of us. Not another strategy paper. Not twenty-seven national AI plans politely contradicting each other. One strategy, run like the defence programme it actually is. Bring the talent home: the European chip designers in California, the researchers training frontier models for other people’s missions, the founders who left because the capital and the compute were somewhere else. Fund sovereign compute the way we fund motorways. Build open-source models that belong to us, not fine-tuned copies of weights that an American or Chinese lab decided, this quarter, to let us borrow. And use the leverage we already hold at the bottom of the stack instead of apologising for having it.

None of this is comfortable. All of it is cheaper than the alternative. The alternative is the world we watched yesterday, made permanent. A Europe that builds elegant applications on top of an intelligence it does not own, cannot govern and is not allowed to keep, running on machines it is forbidden to buy, available right up until the morning someone in another capital decides otherwise.

We own the stack, or we keep renting our future from people who have just shown us, in writing, that they will switch it off the moment it suits them.

A note on the facts

Everything in this piece is real, including the part that reads like a thriller.

• The 12 June shutdown happened. Claude Fable 5 and Claude Mythos 5 became available on 9 June 2026. Three days later, on 12 June, the US government directed Anthropic to suspend access to both for any foreign national, whether inside or outside the United States, including Anthropic’s own foreign-national employees. Unable to filter its users by nationality, the company disabled both models for everyone, worldwide, three days after release. The order came through national-security export authority; the letter gave no detailed justification, and Anthropic said it disagreed with the action. (Statement on the US government directive to suspend access to Fable 5 and Mythos 5, Anthropic; US orders Anthropic to disable AI models for all foreign nationals, Al Jazeera; Anthropic Pulls Its Most Powerful AI Models After U.S. Bars Foreign Access, Time)
• The frontier is American and increasingly governed by export law. The US already restricts the sale of its most advanced AI chips through the “AI Diffusion Rule” and its successors, a tiered regime that blocks flagship parts like the H100, H200 and Blackwell from most of the world by default. The same export-control machinery governs both chips and the software that runs on them. (United States export controls on AI chips and semiconductors, Wikipedia; AI Export Controls: Navigating Chip Restrictions Globally, Introl)
• Nvidia designs the overwhelming majority of the world’s AI silicon. Estimates cluster around 80 to 90% of the AI-accelerator market: Bloomberg data put Nvidia at roughly 86% of AI data-center revenue in late 2025, and the company held about 92% of discrete GPUs in the first half of 2025. (Charted: The Battle for AI Data Center Revenue, Visual Capitalist; Data center / AI chip revenue of Nvidia, AMD & Intel 2025, Statista)
• The narrowest chokepoint in the supply chain is Dutch. ASML, in Veldhoven, is the world’s sole supplier of EUV lithography machines, the only tools that can print chips at 7 nm and below. Every advanced chip on Earth, including every Nvidia part the US rations, depends on a machine exactly one company knows how to build. (ASML Has a Monopoly on the Most Important Machine in Tech, The Motley Fool; ASML: The Key Bottleneck In The Global Semiconductor Supply Chain, Seeking Alpha)
• Europe does design its own AI silicon. Axelera AI, based in Eindhoven, builds energy-efficient AI inference chips. Its Europa AIPU launched in October 2025, the company has raised more than 450 million dollars (including a 250 million round in 2026), and its Titania chiplet is funded by the EU’s EuroHPC programme as part of DARE, the open-source RISC-V sovereignty initiative. The caveat: Axelera targets inference at the edge, not the large-scale training where Nvidia dominates. (Axelera AI, Wikipedia; Dutch AI inference chipmaker Axelera AI raises $250M, Tech.eu; Axelera AI Secures up to €61.6 Million Grant, Axelera)
• Europe answers with twenty-seven plans, not one. As of December 2024, 24 of the 27 EU member states had adopted national AI strategies, and the OECD itself flags the result as fragmented policy with limited cross-border coordination. (Progress in Implementing the EU Coordinated Plan on Artificial Intelligence, OECD)