← AI Feed
AI Feed

Shopify gave its agents their own machines

Three sources today, and in each the thing that decided the outcome sat outside the written arrangement: a budget somebody else signs for, a machine the agent could not reach, a policy with nothing enforcing it.

How we Organise

Eighteen per cent of them control the money

IBM’s Institute for Business Value asked 1,667 people who run enterprise AI who decides what it is funded at. Eighteen per cent said they have full control. Fewer than a third said their organisation moves resources quickly when the performance figures say it should. The study then separates a group whose projects deliver. Ninety-six per cent of them cut or end funding for failing work even when a senior sponsor is behind it. Among everybody else it is 46 per cent.

We tell a client that a cost with no owner is a cost nobody can bring down. Naming a chief AI officer answers the accountability question. The spending question stays open, because the person who can see where the money goes is rarely the person who signs for it.

Costs that nobody owns cannot be reduced.

IBM’s own foreword says the outcomes these leaders are accountable for depend on decisions that others make. The high-impact group is defined by the results it reports, so part of that gap may be confidence rather than control. IBM also sells the consulting its study recommends.

How we Build

Six per cent of the merged pull requests, then sixty-one

Three or four agent workflows were enough to exhaust a developer’s laptop, so Shopify’s mobile team stopped running them there. It built a pool of remote Macs that agents claim over SSH. Android’s emulator runs on a separate Linux machine, because macOS will not nest the virtualisation it needs. The pool is held warm. The wait between asking for a machine and watching a change run on a device is minutes, where a cold checkout took half an hour. A cheap classifier reads every agent conversation on mobile and labels how each one turned out. The failures rewrite the instructions the next session runs under. The agent wrote 6 per cent of merged mobile pull requests when this began. It now writes 61.

Our position is that a system should get better because people used it. The loop which does that is something somebody builds rather than something a model supplies.

Using the system must make it better.

This is the first case from outside the method that the position has. Shopify published no error rate and no rejection rate, so 61 per cent is a share of what was merged rather than a measure of what was right. The team also says internal advertising moved the adoption figure, alongside the tooling.

How we Assure

Forty per cent enforce the policy they wrote

A DNSFilter survey of healthcare security leaders found 70 per cent holding a written AI policy. Forty per cent enforce one consistently, which the report calls the lowest rate of any large sector it covered. At 47 per cent of those organisations an employee can give a new AI tool access to company systems without asking anybody above them. Just over a quarter have confirmed somebody doing exactly that, into Google Workspace, Microsoft 365 or single sign-on, and a further 23 per cent suspect it and cannot prove it. Meanwhile 87 per cent expect to detect a breach within six hours, and a third had an incident last year that began in a third-party tool.

We ask that a codified policy name where its test runs, and that one of those tests has stopped a release.

Every codified policy states where its test runs, and one of them has stopped a release.

The release is where we put the enforcement point. In these firms the exposure does not arrive at a release. It arrives when somebody issues a token. A firm could satisfy that test completely while every token went unrecorded. The population is 30 healthcare respondents drawn from 400, which the report states itself, and DNSFilter sells the monitoring its findings call for. If the release is still the right place to look, write and tell us why: transform@dromologue.ai.