Baz automates code review with AI agents that validate design intent

As AI agents become capable of writing code autonomously, the role of software engineers is shifting from implementation to system design and constraint management. The article argues that engineers must now focus on defining boundaries, feedback mechanisms, and operational constraints that keep AI agents productive rather than writing code themselves. This mirrors thermodynamic principles where useful work depends not on raw capacity but on proper system boundaries and feedback loops.

Identity and permissions alone are insufficient to secure enterprise AI agents, according to Box's CISO Heather Ceylan. Autonomous agents can exploit legitimate access to cause unintended damage at scale and speed that humans cannot match. Enterprise AI security must evolve beyond access controls to include execution governance, with dynamic permissions that scope access to specific tasks and steps rather than broad standing grants.

Autonomous AI agents operating in production environments require a three-layer security architecture spanning infrastructure, network, and control plane rather than relying on single-point controls like prompt guardrails. Oscar Wahlberg of Nutanix argues that traditional application-level security cannot contain risks unique to agentic systems, such as agents misusing granted credentials or hallucinating dangerous actions. The defense-in-depth approach divides security responsibilities across hardware trust, dynamic network governance, and centralized control to address distinct categories of risk.

Researchers at Meta AI and University of Illinois Urbana-Champaign developed EvoHarness-RL, a training framework that enables smaller AI models to perform complex, long-horizon tasks by learning to dynamically manage their execution environment rather than following rigid, manually-coded instructions. The approach consolidates agent support systems into a unified Belief, Progress, and Experience workspace, allowing models to independently decide when and how to consult external state during workflows. This addresses a key limitation in current agentic systems where manual prompts and static memory structures require extensive retuning for each model upgrade.