THE SECRET TO SUCCESSFUL AI AGENTS ISN'T THE CODE

A capital-allocation exercise across 150 cities used to take Uber 15 hours. It now takes 30 minutes.

Financial pacing reports took two days. Now they take ten minutes.

Web QA went from two weeks to 50 minutes.

How?

Uber paired around 30 of its most AI-proficient engineers with domain experts from finance, legal, HR and other business functions. Each pod got two weeks to get the work done.

And the sprint schedule offers a glimpse into what makes effective agentic sprints.

Days one and two: shadow the expert. Watch every step. Find the judgements, exceptions and workarounds missing from the process diagram.

Day three: rank the opportunities by scale, repetition, business impact and data availability.

Days four and five: build a working agent beside the person doing the job.

Days six to nine: test it with other people doing the same work. Does it generalise? Does it actually make the job better?

Day ten: ship.

Sixteen pods. Sixteen functions. Two months.

Uber’s CTO said process diagrams and documentation were not enough: ‘You have to understand how the work actually gets done.’

And the biggest wins came when the pods stopped automating isolated tasks and redesigned the workflow itself – removing handoffs, cutting unnecessary approvals and replacing legacy tools.

The workflow becomes the unit of automation.

That time can be booked as a cost saving. Or, better, turned into capacity for work that never fitted into the day.

Two days observing. One choosing. Two building. Four validating. One shipping.

In this new agentic world of ours, writing code isn’t the bottleneck. Success comes from understanding the work.

Or, to put it a slightly different way, agentic development has to happen where the work happens.