What are the main challenges deploying autonomous AI agents?

Zibi Braniecki

Member of Technical Staff · Perplexity

Unpredictability

Zibi Braniecki says the challenge is unpredictability: agents may do something you didn't expect, not from malice but just trying to fulfill your task.

The answer starts at 1:39.

JS

Jim Scharf

Chief Technology Officer · Mongodb

Governance, Security, Cost

Jim Scharf says the challenges of taking agents to production include how to manage, govern, deploy and version them, and specifically securing, governing, auditing, and managing cost, which he says peers are wrestling with.

The answer starts at 21:51.

Rama Akkiraju

VP, Enterprise AI · Nvidia

Observability

Rama Akkiraju says observability is key, especially to understand agent behavior before certifying an agent.

Roberto Milev

Chief Architect · Navan

Cost Predictability

Roberto Milev says cost predictability is a challenge because it is very hard to predict and manage agent costs, requiring guardrails and fallbacks such as cheaper models for some tasks.

Peter Desantis

SVP, AWS Utility Computing Products · Amazon Web Services (AWS)

Infrastructure and Systems

Peter Desantis says the challenge is infrastructure and systems, arguing AI infrastructure is not a chip or model problem but a systems problem as AI agents become more autonomous.

Konstantin Klyagin

Founder · Redwerk

Auditability

Konstantin Klyagin says auditability is a challenge, requiring testing whether an institution can reconstruct what happened and when; if you can reconstruct it, you own your solution, and if not, it owns you.

Daniel Herrington

Chief AI Officer · The Zebra

Code Review Bottlenecks

Daniel Herrington says one of his biggest bottlenecks is how to do code reviews on sensitive code when his team can generate code so fast. He calls sensitive code reviews a hard sticky point for them.