AI Agents Face Speed Bumps Due to Old Systems, Say Tech Leaders

At a recent technology conference, leaders from LinkedIn, Walmart, and Zendesk shared a common finding: outdated infrastructure, not AI models, is slowing down AI agents. These agents are designed to work quickly, but the systems they run on were built for human-paced tasks. Animesh Singh from LinkedIn explained that tools like Kubernetes take seconds to start, which is too slow for agents. His team fixed this by pre-loading containers. They also found that letting AI evaluate itself led to errors, so now about 80% of tasks use fixed code, with AI only for reasoning. Desiree Gosby from Walmart said their agent tool became so popular that employees built too many similar agents. They solved this by creating rules to find the best versions and share them without slowing down engineering. Sami Ghoche from Zendesk noted that having billions of customer conversations doesn't help if data pipelines aren't ready. He stressed investing in data systems instead of just feeding everything to a large AI model. All three agreed on keeping flexibility. They build systems that can switch between different AI providers. Singh advised building for independence, so companies aren't stuck with one model. Ghoche said most businesses want to own their own models but still use top labs for advanced reasoning. The leaders also gave simple advice: start with good evaluation tests, give employees access to AI tools early, and build systems that can change with new technology. Their main message is that fixing old infrastructure is key to making AI work fast.