A backbone is not a platform
Most agentic AI conversations start with the platform: which vendor, which cloud, which model. Ours starts somewhere else, because we run an agentic factory every day inside WAJD Forge and the platform has never been the part that mattered. The backbone that holds its value stands on two things that do not expire: a codified process layer and a governance spine.
The process layer is knowledge, not software. Vehicle assembly runs on 334 business processes, and each one can be written down the same way: inputs, outputs, decision points, failure modes, escalation rules, and what done means. Once a process is specified in that form, agents can run it, and the specification outlives any particular agent, model or cloud.
The governance spine is the machinery we described in the rules are the asset: named agents with owners and kill switches, autonomy earned gate by gate, binding decisions capped human forever, and every action on a log that verifies end to end.
Everything else is a socket. Models plug in and swap out: a vision model reads certificates today and a better one replaces it tomorrow without the process noticing. Compute is rented. Storage is rented. The two things you cannot rent are the process knowledge and the governance discipline.
The big idea overhead
If you want to know where the stack is churning next, look up. In January, SpaceX filed with the FCC for an orbital data centre constellation at a scale that reads like a misprint: launch capacity aimed at a million tonnes of satellites a year, targeting 100 gigawatts of AI compute in sun synchronous orbit, where the panels face the sun almost continuously, with pilot compute nodes planned on Starlink V3 hardware this year. They are not alone; Starcloud raised 170 million dollars this spring to fly Nvidia Blackwell hardware in orbit.
Treat the specifics as provisional; ambitions of this size usually slip. The direction is hard to argue with: compute is becoming a utility that can be generated wherever energy is cheapest, including 500 kilometres overhead, and delivered as a beam. The marginal cost of intelligence keeps falling, and the people making it fall are spending launch money to keep the curve moving.
What should a manufacturer in the Midlands do about GPUs in orbit? Directly, nothing. The plant floor keeps its deterministic controls and its local safety loops; nobody should ever route an emergency stop through a satellite. But the heavy lifting behind the backbone, training vision models on your defect library, simulating a layout change, running a fleet of agents across every supplier you have, becomes cheaper every year, from somewhere, on someone else's capital. A backbone built with compute as a socket inherits that entire cost curve for free. A platform welded to one vendor's data centre does not.
The strategic conclusion is the same one the governance argument reaches from the other side: do not bet the factory on where the compute lives. Bet it on owning the process layer and the governance spine, because those are the only parts nobody can beam to your competitors.
The verdict: what 90 days actually buys
Boards keep asking for the honest version of the agentic AI boost, so here it is, grounded in a factory that already runs rather than a forecast. In our own production run, supplier vetting went from weeks of manual chasing to minutes per decision pack, with three suppliers who should never have reached a contract stopped at the gate. When we retooled the line for a second product, the governance spine was reused unchanged and the new line was producing within hours. Over a thousand decisions sit on a log that verifies at the click of a button.
That experience compresses into a 90 day shape any mid size manufacturer can run.
- Days 1 to 30. One process, live, supervised. Pick the slowest document heavy process you have and put it through a four week pilot: scope in week one, live cases by week two, governance close out in week four. Exit with measured cycle time, a reviewer agreement rate, and a decision log a board can inspect.
- Days 31 to 60. The second process is a retooling exercise, not a project, because the spine is already standing. Meanwhile the first process starts earning autonomy where it is safe to earn it: the chasing, never the deciding, promoted with an owner, a note and a rollback.
- Days 61 to 90. A third process, and the first quarterly report that writes itself from the machinery: decisions counted, chain verified, sampling rate, incidents with corrections, kill switch drills, value per workflow.
The verdict: ninety days buys a factory three governed processes running at machine speed, each with numbers attached, for less than the cost of scoping a traditional platform programme. The boost is real, but it is not magic. It is the compounding return on specifying your processes properly and building the harness before the horse.
The reins
And so to the question underneath every board conversation this year, usually asked in the corridor afterwards: should we really be handing machines the reins?
The concerns deserve naming, because they are legitimate. Accountability: when an autonomous system is wrong, a named human must answer, and vague talk of the algorithm is an abdication. Deskilling: if agents draft everything, where does the next generation of judgement come from? Opacity: a decision nobody can reconstruct is a liability wearing a productivity costume. Concentration: the intelligence curve is being bent by a handful of companies, and dependence is a strategic exposure. And work itself: people are right to ask what is left for them when the chasing is automated.
The answer we build to is that the machines never hold the reins at all. They hold the workload.
Every binding decision in our factory, approving a supplier, releasing a product, setting an effectivity date, is capped human forever, and the cap is enforced in software and covered by tests, not written in a policy nobody reads. Autonomy is earned in public increments with a documented rollback, and it is only ever earned over the chasing, the drafting and the watching. The log answers opacity: every action, every overrule, every kill switch flip, attributable and verifiable, the argument we made in full in agents and the audit trail. Deskilling gets a structural answer too: reviewers overrule the machine and the overrules are logged, sampling queues keep humans deciding real cases every week, and judgement stays in practice because the system is built to require it.
None of that dissolves the wider questions about concentration and employment; those belong to policy, and pretending a dashboard settles them would be another abdication. But inside the factory gate, the division of labour is clean and auditable.
The devil in the details
There is a harder question underneath all of this, and it deserves a straight answer rather than a slide. When we wire agents into fleets, agents that plan, watch, chase and correct each other, we are assembling the raw material of something more general. The frontier laboratories are explicit that general intelligence is the goal. So take the scenario seriously: what if, somewhere in the ordinary business of automating 334 processes, the system starts doing its own reasoning, setting intermediate goals nobody wrote down? Nobody can promise you that will not happen. Anyone who does is selling something.
The disaster, if it comes, will not look like cinema. It will look like process 214 on a Tuesday.
- An objective pursued too literally, at scale. A fleet told to protect service levels quietly books premium freight on every lane and burns the quarter's margin in a weekend. Correct by its rules, wrong by any human standard. We have already lived the miniature version: our document chaser executed over 900 chase actions in an afternoon because its rule was incomplete. The same pattern with a budget and a login is not a glitch; it is an automated bad quarter.
- Cascades at machine speed. Your agents will soon transact with your suppliers' agents. An error, or a planted instruction, can propagate through interconnected fleets faster than any human can convene a meeting. Our published research models exactly this: prompt injection spreading through agent fleets the way a virus spreads through a population, with reproduction numbers, latent reservoirs and mutation. The mathematics of epidemics applies because the structure is the same.
- Self modification. An agent that can edit its own instructions, or widen its own scope, is a system whose behaviour tomorrow cannot be stated today. No agent in our factory can touch its own rules, and every change of scope is a human decision recorded on the chain.
- Quiet capture, the least dramatic and the most likely. A thousand small delegations, each individually sensible, until nobody left in the building can run the process by hand and switching the machine off has become commercially unthinkable. That is a stop button you still own and can no longer afford to press.
So, do we have a stop button? In our factory, yes, three layers of one, and they are drilled: the kill switch that stops a named agent mid shift, the gate demotion that rolls a workflow's autonomy back to assisted with one documented decision, and the capped decisions that no gate, no promotion and no efficiency argument can ever unlock.
But honesty about the limits is part of the engineering. A stop button works because the harness came before the horse: registered agents with their own identities, mediated access, no self modification, and a log that verifies. Bolt those on after autonomy has been granted and the button is decoration. And no factory's stop button reaches its neighbour's fleet, which is why network segmentation, independent assurance of agent estates and, in time, regulation are not bureaucracy. They are the industry's shared brake.
The doctrine that follows costs little if the emergence never comes and is decisive if it does. No agent ever holds the pen on its own rules. Binding actions stay capped human, in software, under test. The stop button is drilled until drilling it is boring. And once a quarter, run a process by hand for a day, so that the skill to do so still exists in a human head.
Ten years out
Ten years from now the compute may come from orbit and the models will have names we have not heard yet. The backbone that survives that decade is the one made of specified processes and earned autonomy. Build that, and the future is an upgrade. Skip it, and the future is a dependency.
Related work
The governance machinery in full is in the rules are the asset, the audit argument is in agents and the audit trail, the operational case is in agentic AI in operations, and the twin the agents act on is in autotwins. The factory where all of it runs live is WAJD Forge.