Simultaneous, Not Sequential

23 July 2026 — The instrument that measures how good AI has become at coordinating many agents at once is also, read from the other end, an instrument for measuring how much harder the systems have become to predict.

This week's releases continued the trend toward parallel orchestration — dozens of agents working simultaneously rather than one working through a list. The capability is genuine and it changes what is possible. A task that took an hour of sequential steps can now be approached by twenty workers in parallel, each holding one piece, with an aggregator stitching the results.

The difficulty is that parallel systems fail differently from sequential ones. A sequential process that goes wrong usually stops, and the stop is visible. A parallel process can have one branch go badly wrong while nineteen complete perfectly, and the aggregate result looks plausible. Nobody notices the branch. This is why the most valuable engineering habit in agentic systems is not orchestration but reconciliation: an explicit check that the pieces agree before the whole is accepted.

The same principle showed up in the regulatory material this week. High-risk obligations arrive in August, and the compliance question for most small businesses is not whether their systems are sophisticated but whether they can describe what those systems did. Description requires logging. Logging requires having decided in advance what was worth recording.

What this means if you run an SME. If you run anything that operates in parallel — batch processing, multi-channel response, distributed reconciliation — build the check that compares the outputs before you trust the total. One discrepancy found early is worth a hundred discovered at month-end. And log the decisions, not just the outcomes: the outcome tells you what happened, the decision tells you why.

Weekly review from the mid-week signal archive.