Faster AI coding stalls when decisions, sign-offs and context stay slow

Writing code is no longer the slow step at GEICO. Ford Prior, a senior staff engineer there, said on the AI Engineer show that the delay now sits in everything around the code: decisions, sign-offs and context. His answer was to rebuild that surrounding process rather than speed up the editor further.

What was said

Prior borrowed Gene Kim's terms. The inner loop is coding and building. The outer loop is everything else. He said the build step at his workplace has been largely wiped out as a time cost, yet work still moves slowly.

He gave a number. About six months ago, GEICO checked its backlogs. After patches and minor upgrades, engineers had 25% of their time or less, on a good day, for major changes. So the company put AI agents on the patches and minor work.

Major work stayed stuck. Prior said, "we find ourselves rate limited by our ability to collaborate in rooms full of the right individuals summoned at the right time." He added that teams also lacked a way to pull facts and code analysis together as questions came up.

His fix was an AI-native war room. It has decision makers and subject experts, one driver who operates the costly AI tools, and agents wired into chat, design tools and code. His bet: $1,000 in tokens and one day could map one legacy claims domain and propose one real change. That change was an API that AI agents can call. He ran the test three times. In the rentals run, he said, the agent found a link that would save 8,500 calls a month.

Why it matters

Our reading: a team can buy faster coding and still ship at the old pace. If approvals and missing context set the speed, a coding-speed metric will mislead managers. Prior's example suggests tracking how long decisions wait.

It also shows where skill goes. In his war room, people supply judgment: choosing, approving, questioning. One person drives the tools. Prior said the deep context built about the legacy system is built once and can be reused. His advice was blunt: "find some group in your organization that doesn't know how to make decisions. Deploy this. Drive it. Control the cost."

The other side

This is one company, told by someone whose job includes sharing such stories to attract engineers. The excerpts give a result only for the rentals run. They do not show whether the damages and salvage runs shipped a change in a day. Prior also said an agent found a production bug, but he could not share details.

The war room still needs the right people in one room, which was the original bottleneck. It shortens that wait to a day rather than removing it. Cost is a real limit too. Prior warned that letting an agent schedule its own check-ins can start a new paid session each time. He did not claim every company will win this way.