London — In my agent workflows, I check the work before accepting the result. That means defining the deliverable and its checks before a task starts, reviewing the output, and keeping a human approval step for consequential decisions. These controls help me find errors; they do not make generated work error-free.
What does verification at volume require?
I identify task dependencies before running agents in parallel, agree the interfaces between their changes, and attach verifier commands to each task. I reject reports that lack the declared deliverable. Security and compliance work also gets an independent review that challenges the first finding, followed by human approval.
A failed check needs a diagnosis. I use the MAST failure categories to distinguish problems in the task specification, coordination between agents, and verification. That helps identify whether to change the instructions, the workflow or the checks.
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 in this dispatch tests that approach — 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.
These checks also take time and tokens. The question I keep returning to is which checks catch useful errors and which add overhead. More agents or more review steps are not, by themselves, evidence of better results. That needs measurement on the work the system is meant to do.
— V. U., 22 Aug 2026