A grid-scale battery already runs itself. Its energy-management system reads state of charge and availability from the battery’s own management system and dispatches on that, every few seconds, on economics no one checks by hand. Now suppose one pack’s management system has been quietly misreporting since a firmware update three weeks ago. The optimiser acts on the number it is handed: fast, confident and wrong, and wrong again on the next cycle. No agent required. That is today.
The autonomous era does not introduce this. It deepens it. Connected critical infrastructure has a structural blind spot: the devices were never built to be observed from outside the vendor that made them, so everything above them reasons on what they report. Autonomy does not close that gap. It reasons across it faster, and acts on it more confidently. This piece is about the direction connected infrastructure is travelling, and how the forces meant to improve it, autonomy, encryption, ever more connection, are already compounding the blind spot today.
The blind spot, briefly
Failures live in the gaps between devices, vendors and teams, not inside any one of them, because the devices are opaque and everything above them infers. Each team sees its own layer; each tool sees what it was built to see; no one sees the whole. This is not a knowledge problem that better people or more dashboards fix. It is how the category is built, and it is why network inference, vendor dashboards and predictive models running on assumed-good telemetry do not close it.
What autonomy promises
The pitch is specific: predictive maintenance that catches failure early, diagnosis compressed from days to minutes, orchestration that coordinates across vendor boundaries humans cannot reach. On clean, well-instrumented inputs some of it is real, and the promises are not, as a class, dishonest. But there is a difference between what autonomy is and how it is being sold. Most of what is shipping is AI features bolted onto tools that already could not see the device layer. Faster reasoning above the gap does not close the gap; it runs the same distance from the source at higher speed. And the further the reasoning sits from where behaviour actually happens, the more it rests on representations of representations, each losing a little fidelity, so an agent can be confidently wrong in proportion to its distance from the device.
What it inherits
The causes of the blind spot sit below the layer autonomy operates on, so replacing people with agents at the top changes none of them.
The opacity is economic and legal, not cognitive. A vendor’s agent sees only what that vendor’s systems expose, and those surfaces are built for the vendor’s purposes, not for whole-system reasoning. Vendors protect their IP; operators do not hand cross-vendor agents privileged access. Agents inherit every one of those limits directly.
And the estate itself does not change. Connected infrastructure is full of cheap sensors with thin telemetry and long-lived industrial kit chosen on supply and price. An agent does not transform the firmware of a 2012 medical device, a 2005 PLC or a £200 temperature sensor. The medical device deployed in 2018 is still running; the battery management system from 2019 will run into the late 2030s. Regulation like SBOMs and the Cyber Resilience Act will help, slowly, and only for kit designed now. The autonomous era will be operating on this installed base long before any remedy reaches it.
What it reshapes
Here is the part the discussion has not caught up with. The blind spot is not only inherited; it is being hardened into the architecture, by the same security direction the era depends on. Encryption between device and cloud is baseline now; hardware-rooted trust, signed firmware, attested boot and confidential computing are becoming procurement defaults in regulated sectors. This is welcome, and largely required by the same rules operators must meet. It also reduces what any observer above the device can see by passive means, and raises the privilege needed to interact actively. So the gap does not close. It moves into a shape that is harder to see into, and autonomy will be expected to work inside that envelope, on a picture it can verify even less than before. None of this is a future constraint; it is already tightening, on kit being specified this year.
What it amplifies
Some of it autonomy makes worse, and you can see the shape of it already, in the automated layer that runs sites today: it acts on what the device reports, fast, and no one checks in the moment. Agents do not start this. They scale it.
A human operator hesitates when something is unclear. A system tuned to act on confident outputs does not, unless that hesitation is deliberately designed in, and the commercial pressure runs the other way. Agents are also good at producing outputs that look right. Acting on incomplete inputs, they generate decisions and reports that read as authoritative because the estate as a whole keeps functioning. The vendor-dashboard problem, a component reporting itself healthy while the integration fails, now extends to the orchestration layer: the agent reports the estate is within parameters because, from its bounded view, it is. The blind spot has started writing confident reports about itself, at a speed no human review cycle can keep up with.
When it does go wrong, it is harder to reconstruct. A post-incident review needs to know what the device actually did. If the device layer is opaque, the trail shows only what the agent decided based on what it inferred. When the agent’s own reasoning is opaque too, as it is for the machine-learning systems now proposed for these roles, the trail is doubly short: you can know the agent acted, not fully why, and not what the device beneath it did. Accountability gets structurally weaker exactly as the systems get faster.
And something quieter erodes. Experienced operators carry judgement that is in no document: where a system is fragile, what an odd reading means, when something is simply not right. It is built over years at the controls, and it is one of the ways people have covered for the blind spot all along. As autonomy takes the routine work, the hours that build that instinct go with it, and without a better picture of the device layer underneath, agents cannot substitute for what is lost.
What has to sit underneath
The argument is structural, so the answer has to be too. This is not the place for a product; it is the place to say what anything the autonomous era stands on, including anything we build, has to satisfy.
It cannot be a vendor’s own agent. Vendor opacity is not a telemetry accident; it is a property of the commercial position the vendor holds, and a foundation that depends on vendor self-reporting inherits the vendor’s interests along with its data. It has to sit outside the vendors whose kit it observes: an honest broker by structural design, not by claim.
It has to sit close to where behaviour originates, so the reasoning above acts on something that has not been compounded through many layers of vendor-shaped inference. The exact form will vary and evolve; the principle does not. Closer to the source than the agents reasoning above it.
Its audit trail has to reach the device, not stop at the agent. When something fails, the question is not what the agent decided but what the device actually did, and how that compares to what the agent acted on. A trail that stops at the agent stops short of accountability; one that stops at the device’s own self-report stops short of the truth.
And it has to be built for an encrypted world, not against one. The same cryptography that blinds passive observation also produces attestation and signed lineage, which prove where data came from, though not that it reflects what the device was doing. A foundation fit for this era works with that direction, or it does not work for long.
Last, its own access has to be earned. A system that asks for deeper integration than it has earned has become part of the attack surface before it has become part of the assurance. The depth it takes, point-in-time, intermittent, continuous, active, must be earned in that order, by demonstrated behaviour. Anything that one day runs without an operator should get there the way everything else in this piece earns its place: last, and on evidence.
The direction is set, and it is already compounding. Something has to sit underneath it, and it will be built either way. The only question is whether it is built before or after the cost of not having it is paid.
Closing
Autonomy is only as good as what it stands on: a picture close to the device, or a construction assembled many layers away from it.
One thing does not get easier as more intelligence is stacked on top: independently establishing what a device actually did. It gets harder, and it matters more, for a diagnosis, for a dispute, for the audit trail, for trusting the agent above. That ground widens as the estate gets more autonomous, not narrower. Autonomy does not lessen the need for evidence from the device. It raises it.