London — The industry is still optimising the wrong half of the problem. Generation is solved well enough to be boring: any competent team can get a model to produce something plausible. What nobody has solved is verification at volume— getting a fleet of agents to produce work you would put your own name on, at a cost that survives contact with a budget, under conditions where being wrong has a consequence.
What does verification at volume require?
The distinction is not academic. It changes what you build. If you believe generation is the constraint, you invest in prompts, models, and context windows. If you believe verification is the constraint, you invest in the unglamorous scaffolding: dependency-aware dispatch so agents do not race each other, frozen interfaces so parallel work composes, per-task verifier commands, structural rejection of any agent report that arrives without its declared deliverable. On security and compliance work, an independent verifier whose only job is to refutethe first agent’s finding, and a human gate that cannot be skipped.
Most of that is not AI engineering at all. It is ordinary distributed-systems discipline, applied to workers that are unreliable in an unfamiliar way — confidently wrong rather than crashed. The failure taxonomy is the tell: agent postmortems here run against MAST, which sorts failures into specification, inter-agent misalignment, and verification. Only the first is a prompting problem. The other two are design problems, and they are where the time goes.
What changes in a regulated operating context?
A regulated operating context makes human ownership, review gates, and audit trails non-negotiable before deployment. That is an internal governance and system-design choice at Altery, not an AI-specific rule created by Electronic Money Regulations. Vasiliy Uvarov’s EMD Individual standing concerns management of an electronic-money institution generally; it is not an AI qualification. The practical effect on the engineering sequence is still useful: the verification layer gets built before the first deployment rather than after the first incident.
The day-job work this month is a test of exactly that claim — moving the legal and regulatory agent team off a flat shared library and into a routed configuration with explicit verifier–builder pairs and written escalation contracts. With token cost forced back into the routing decision, the open question is how much of that routing still earns its keep. On the bench, Cosmic Copilot is the active project: 1,014 commits across the native app and web app in the trailing 90-day window ended 21 May 2026, with the iOS build now ahead.
If the thesis is wrong, the way it will be wrong is this: models get reliable enough that the scaffolding becomes dead weight, and everyone who built verification infrastructure will have spent two years solving a problem that evaporated. I do not think that happens on the current curve — reliability improves, but the tasks we hand agents grow harder at least as fast, so the verification burden moves rather than disappears. Worth revisiting here in a year, with the benefit of being publicly on the record about it.
— V. U., 22 Aug 2026